Pogbagram

OpenAI and Anthropic simultaneously release major model updates

Early this morning, OpenAI and Anthropic released major model updates, launching GPT-5.3-Codex and Claude Opus 4.6 respectively.

Related reading:

OpenAI claims that GPT-5.3-Codex is its first model to play a key role in its own development process, with the team using an early version to participate in debugging, training, deployment management, and evaluation analysis.

The model broke records in multiple benchmarks, including SWE-Bench Pro, Terminal-Bench 2.0, and OSWorld-Verified. It achieved a score of 77.3% in Terminal-Bench 2.0, a significant improvement over GPT-5.2. In OSWorld-Verified, its accuracy improved to 64.7%, approaching the human average level.

OpenAI emphasizes that GPT-5.3-Codex can not only write code, but also perform complex operations across software, support long-term tasks and real-time interaction, and demonstrate stronger autonomy and intent understanding capabilities in tasks such as building web pages and games.

Around the same time, Anthropic released Claude Opus 4.6, emphasizing inference capabilities, reliability, and long context handling. It introduced a 1M Token context window (Beta) for the first time, achieving a recall rate of 76% in MRCR v2 long text retrieval tests, far exceeding its predecessor.

Opus 4.6 outperforms GPT-5.2 by approximately 144 Elo in GDPval-AA (High Economic Value Knowledge Work Assessment), and also leads in assessments such as Humanity's Last Exam and BrowseComp.

Anthropic also launched the Agent Teams feature, allowing multiple agents to collaborate in parallel, supporting task decomposition, independent contexts, and inter-agent communication. In the official demonstration, 16 Opus 4.6 agents autonomously completed a 100,000-line C compiler within two weeks and successfully compiled the Linux 6.9 kernel.

In productivity scenarios, Anthropic deeply integrates Claude into Excel and PowerPoint, automatically generating consistent presentations based on tables and enabling multi-task collaboration within Claude Cowork.

TechCrunch points out that Opus 4.6's Agent Teams allows ordinary developers to experience the work of "commanding an AI team".

OpenAI emphasizes the high reliability and low variance of GPT-5.3-Codex, making it more suitable for engineering execution and maintenance; Anthropic, on the other hand, emphasizes the high upper bound inference capabilities and long text processing advantages of Opus 4.6, making it more suitable for financial, legal, and complex decision-making scenarios.

Bloomberg reports that the release of Opus 4.6 has impacted the financial data services industry, with the stock prices of many listed companies falling in the short term.

It is worth noting that, according to TechCrunch, OpenAI's GPT-5.3-Codex was originally planned to be released at the same time as Anthropic, but after the latter was released 15 minutes earlier, OpenAI immediately adjusted its release time and quickly launched the update.

The iPhone 17e is rumored to be unveiled this month.

According to Macworld, Apple is expected to release the entry-level iPhone 17e via press release on Thursday, February 19th.

Multiple accessory manufacturers and supply chain sources indicate that the core upgrades to the iPhone 17e focus on charging and connectivity. The new phone will feature MagSafe magnetic wireless charging with a maximum power of 25W, while the previous generation only supported the 7.5W Qi wireless standard.

In addition, the iPhone 17e is expected to be equipped with Apple's next-generation C1X baseband, which Apple claims is twice as fast as its predecessor, the C1, and has higher energy efficiency.

In terms of appearance, MacRumors points out that the overall design of the new phone will continue the iPhone 16e, including a single camera and a notch screen. The rumored "Dynamic Island" design will not be implemented in the e series.

Other reports suggest that Apple may equip the iPhone 17e with a U1 or U2 ultra-wideband chip to support the "precise find" feature of AirTag 2, but this has not yet been confirmed by more authoritative sources.

Judging from the release schedule, Macworld believes that Apple is gradually forming a dual-cycle release strategy of "spring e-series, fall flagship". If this rhythm continues, the entry-level iPhone 18, starting in 2027, may also be launched in the spring at the same time as the 18e.

Hurun China 500 List Released: TSMC Tops the List Again, Xiaomi Enters the Top Ten for the First Time

Yesterday, the Hurun Research Institute released the "2025 Hurun China 500," which lists the top 500 non-state-owned enterprises in China and ranks them according to their corporate value.

• TSMC topped the list for the second consecutive year with an enterprise value of NT$10.5 trillion, an increase of NT$3.5 trillion from last year;
• Tencent ranked second with 5.3 trillion yuan;
• ByteDance ranked third with 3.4 trillion yuan, a growth rate of 109%, making it the fastest-growing company among the top ten.
• Xiaomi entered the top ten for the first time this year, ranking eighth with 1 trillion yuan, mainly benefiting from the rapid expansion of Xiaomi's automotive business and the profit improvement brought about by the high-end mobile phone trend.
• Huawei ranked tenth with 850 billion yuan.

Overall, the total value of the top 500 companies reached 77 trillion yuan this year, an increase of 21 trillion yuan year-on-year, representing a growth rate of 38%.

The semiconductor industry performed best, represented by companies such as TSMC, Cambricon, and MediaTek, becoming the second largest industry on the list, with a growth rate exceeding that of the life and health industry. The industrial products industry remained in first place.

Industries such as chemicals, metals and minerals, and energy also saw significant growth, while the real estate sector continued to decline, and the retail sector was among the top decliners.

In terms of geographical distribution, Beijing, Shanghai, and Shenzhen remain the three cities with the most companies, with 59, 57, and 49 companies respectively.

Taipei, Hong Kong, and Hangzhou followed closely behind. Shanghai saw the largest increase in the number of businesses this year. Suzhou emerged as the most favored manufacturing base, while Shanghai was the top choice for R&D centers.

90% of the companies on the list are publicly listed companies, with an average enterprise value of 153 billion yuan. 95 companies are new entrants, including several semiconductor and consumer electronics companies such as Muxi, Moore Threads, and Inspiron.

Sectors experiencing significant value declines this year include food delivery, e-commerce platforms, and some Taiwanese companies, primarily due to the rise of the mainland's AI industry chain and changes in the competitive landscape. The real estate sector continued to contract, with the number of companies listed decreasing from 68 before the pandemic to 12 this year.

The Hurun Research Institute stated that the Hurun China 500 reflects the structure of the most promising private enterprises for the next decade, and the changes in the list show that China's private economy is accelerating its transformation towards high-quality development.

iOS 26.3 RC reveals M5 Max and M5 Ultra, M5 Pro may be cancelled.

According to MacRumors, the iOS 26.3 RC (release candidate) version that Apple pushed to developers yesterday included information about two unreleased Apple M5 series chips, sparking further speculation about Mac product line plans for this year.

Developer Nicolás Alvarez discovered SoCs codenamed T6051 (H17C) and T6052 (H17D) in the system code. The "17" corresponds to the M5 series, and the letters "C" and "D" continue Apple's existing naming conventions for Max and Ultra chips, respectively.

According to the previous numbering system, the standard M5 is H17G, while the M5 Pro should correspond to H17S. It's worth noting that the T6050 (H17S) designation did not appear in the code this time.

Apple previously used the letter "C" for all Max chips, including M1 Max (H13C), M2 Max (H14C), M3 Max (H15C/H15M) and M4 Max (H16C).

Ultra chips all use "D", such as M1 Ultra (H13D), M2 Ultra (H14D), and M3 Ultra (H15D). No Ultra version was released for the M4 series.

The report points out that, based on this pattern, the two chips appearing in iOS 26.3 RC can almost certainly be identified as the M5 Max and M5 Ultra, respectively.

It's worth noting that the market had widely expected Apple to release a new MacBook Pro equipped with the M5 Pro and M5 Max, but the absence of any indication of the M5 Pro in the code suggests three possible directions for the product line:

• The M5 Pro still exists, but it hasn't been added to the iOS code yet;
• Apple is adjusting its chip naming system;
• The new MacBook Pro may only be available in two configurations: M5 Max and M5 Ultra.

The report indicates that Apple is widely expected to release a new MacBook Pro powered by the high-end M5 chip in the coming weeks. Meanwhile, Apple is also developing a new Mac Studio based on the M5 Ultra, but there is currently no definite release date.

Following its own ingot-shaped red envelopes, WeChat has blocked Baidu Wenxin Assistant's red envelopes.

According to Yicai Global, Baidu's Wenxin Assistant red envelope sharing activity was blocked on WeChat yesterday. When users clicked the relevant links, they saw a message stating that the webpage contained content that induced or misled downloads or redirects, and that it needed to be opened in a third-party browser, similar to the previous restrictions imposed on Tencent's own "Yuanbao Red Envelope".

After being restricted by WeChat, Baidu has changed the activity format to "password red envelopes," requiring users to manually copy the password and participate through the Baidu App. Baidu has not yet responded to this matter.

In an announcement released yesterday, WeChat stated that it had received complaints from users regarding the Yuanbao activity, believing that it induced users to share frequently through methods such as "completing tasks" and "receiving red envelopes," which disrupted the platform ecosystem and affected user experience. Therefore, it took action against the illegal links and restricted them from being opened directly within WeChat.

Interestingly, Baidu Wenxin Assistant's red envelopes were briefly restored and could be opened normally within WeChat. Baidu officials even joked that "they treat Wenxin better than Yuanbao (a type of in-game currency)," but the relevant links were subsequently blocked again.

Furthermore, according to actual testing, the link to the newly launched Qianwen App's 3 billion yuan free order campaign was also blocked by WeChat.

Qianwen App's 3 billion yuan free order campaign officially launched

This morning, Qianwen App officially announced the launch of its "3 Billion Yuan Free Orders During Spring Festival" campaign. This campaign represents Alibaba's largest investment in any Spring Festival promotion to date.

The event page shows that the first wave of free orders will run from February 6th to February 12th. Users can receive a 25 RMB no-minimum-spend free order card after updating the Qianwen APP.

This card is primarily used in the "AI-powered one-sentence ordering" scenario. Users can simply issue voice commands to the AI ​​to make purchases at over 300,000 beverage stores nationwide, covering mainstream brands. Furthermore, this free order benefit can also be used to buy New Year's goods or order takeout through Taobao flash sales.

To enhance its social sharing capabilities, the Qianwen APP has implemented a referral mechanism: for every new user invited to download the app, both the inviter and the recipient receive a 25 RMB discount card, with a maximum of 21 cards per person and a cumulative total of up to 525 RMB. If three new users are successfully invited on the same day, the inviter can also participate in a draw for a "Qianwen AI Life Card" worth 10,000 RMB.

As early as January 15, Qianwen APP had already been integrated with Alibaba's ecosystem scenarios such as Taobao Flash Sale, Alipay, Taobao, Fliggy, and Gaode Map, aiming to connect the entire chain from consumer decision-making and transaction to fulfillment.

It is reported that the second wave of activities will be launched on February 13, at which time cash red envelopes with a maximum value of 2,888 yuan will be available.

Google's parent company goes all in on AI, with annual revenue exceeding $400 billion.

Yesterday, Alphabet, Google's parent company, released its financial results for the fourth quarter and full year of 2025, showing that revenue and profit exceeded expectations across the board.

Meanwhile, Alphabet significantly raised its capital expenditure guidance for this year to between $175 billion and $185 billion, almost double that of last year, making it one of the most aggressive AI infrastructure investors among global tech giants.

Alphabet reported fourth-quarter revenue of $113.828 billion, up 18% year-over-year; net income was $34.455 billion, up 30% year-over-year. Full-year revenue for 2025 is projected to surpass $400 billion for the first time, reaching $402.836 billion, with net income of $132.17 billion.

In its earnings report, Alphabet emphasized that the company is at a critical stage of expanding its AI infrastructure. Capital expenditures this year are projected to reach $175 billion to $185 billion, significantly higher than the market's previous consensus estimate of approximately $120 billion.

Key expenditures include data center expansion, procurement of high-performance chips (such as TPUs), and upgrades to energy infrastructure. CEO Pichai stated that the company remains in a state of "supply constraint" in AI training and inference, and any new computing power, even if deployed, will be quickly consumed.

Google Cloud was the strongest performing segment this quarter, with revenue increasing 48% year-over-year to $17.664 billion, surpassing Microsoft Azure's growth rate and marking the fastest pace in nearly four years. Cloud order backlog grew 55% quarter-over-quarter to $240 billion.

The Gemini ecosystem continues to expand. The Gemini App now boasts over 750 million monthly active users, an increase of 100 million from the previous quarter.

Pichai stated that through model optimization and improved computing power utilization, the unit service cost of Gemini decreased by 78% last year. The launch of Gemini 3 is considered a significant milestone and has been deeply integrated into core businesses such as search, cloud, and advertising.

Advertising revenue maintained steady growth. Google Search and other revenue increased by 17% year-over-year, while YouTube advertising revenue increased by 9%, bringing YouTube's total revenue for the year to over $60 billion. Management noted that AI improved the understanding of user intent, significantly enhancing the monetization capabilities of ads in complex search scenarios.

Other key points from the financial report are summarized below:

• Fourth-quarter operating profit was $35.9 billion, with an operating margin of 31.6%, including a one-time compensation charge of $2.1 billion from Waymo; excluding this, the profit margin would have been higher than market expectations.
• Research and development expenses increased by 42% year-on-year, mainly driven by AI-related investments;
• Other bets revenue was $370 million, down 7.5% year-over-year; Waymo recorded $2.1 billion in share-based compensation expenses due to financing, dragging down overall profits.
• Traffic acquisition cost (TAC) was $16.597 billion, up 12% year-over-year;
• Google's services revenue was $95.862 billion, up 14% year-over-year, and remains the company's primary source of revenue.
• YouTube's advertising revenue reached $11.383 billion, a 9% year-over-year increase; combined subscription and advertising revenue surpassed $60 billion for the year.
• Google Cloud's annualized revenue has exceeded $70 billion, with AI demand driving growth in large enterprise orders;
• Gemini's penetration rate in the enterprise market is increasing, with 95% of the world's top 20 SaaS companies now using Gemini technology.

Anthropic's Super Bowl ad mocks ChatGPT; Ultraman: Funny but dishonest.

According to The Wall Street Journal and Business Insider, Anthropic will run its first ad during this year's Super Bowl, directly and playfully targeting its competitor OpenAI.

The report shows that Anthropic created multiple ads for its AI chatbot Claude, with the core theme of "intrusive ads should not appear in AI conversations".

One 30-second ad, set in a fitness scene, suddenly inserts a sales pitch for shoe insoles into a normal conversation, simulating an advertising experience users might encounter on ChatGPT. The ad concludes with the tagline, "AI is adding ads, except for Claude."

Although the advertisement did not name OpenAI, its target was obvious.

Previously, OpenAI had announced that it would test ads in ChatGPT and run its own ads during the Super Bowl.

OpenAI CEO Sam Altman responded on X, saying he found the ads "funny" but also "dishonest," emphasizing that OpenAI would not insert promotional content in the way the ads were presented, and stating that Anthropic's portrayal was misleading.

He also pointed out that OpenAI's goal is to make AI available to more users for free, while Anthropic's products are priced higher and have a narrower audience.

According to reports, US tech companies spent $333.6 million on AI-related TV ads last year, a 43% year-on-year increase; digital advertising spending exceeded $426 million. A 30-second Super Bowl ad slot this year cost over $8 million, excluding production costs.

Anthropic stated that it will insist on keeping Claude ad-free, believing that adding ads to chat products would undermine user trust and could lead to an expansion of advertising due to business pressures.

Company co-founder Daniela Amodei stated that adding ads to scenarios where users frequently share personal and medical information is "exploitative." However, Anthropic has not completely ruled out the possibility of exploring advertising models outside of the chat experience in the future.

Dreame CEO Yu Hao responds to controversy over high profile: Times have changed, and individuality and diversity should be encouraged.

Recently, Yu Hao, founder and CEO of Dreame Technology, has sparked widespread discussion due to his numerous high-profile remarks.

At a recent farewell party, he said that he would become the world's richest man in a few years, and that he would then tell Sa Beining, "I have no interest in money," at Versailles.

Previously, he had publicly stated that "DreamEco will become the first trillion-dollar corporate ecosystem in human history."

These remarks sparked controversy online, with many netizens questioning their overly flamboyant behavior and even worrying that the company might "run away with the money."

In response to the criticism, Yu Hao posted an article yesterday, emphasizing that he had already considered the "risks of being high-profile," but believed that the current social environment is different from the past agricultural society, and the "low profile and moderation" encouraged by traditional culture are not entirely applicable to the present.

He pointed out that the low productivity of agricultural societies made following fixed rules the most efficient solution, thus discouraging individual expression;

In modern society, more developed regions tend to encourage individuality, diversity, and openness. He stated that pursuing this requires embracing a more open and inclusive culture, and being willing to "stand out" for values ​​rather than interests.

In his response, Yu Hao reiterated that the mission of the mission is to "use technology to promote the progress of human society and civilization," and said that he would "defend this mission with his life."

RAM prices surged 3–4 times, Steam Machine release delayed, pricing forced to be recalculated.

According to The Verge, the Steam Machine, Steam Frame, and Steam Controller, originally scheduled for release in the first half of this year, will be delayed due to the continued tight supply of global memory and storage, which has led to a significant increase in the price of key components. Valve needs to reassess pricing and shipping timelines.

Valve stated that the team originally planned to announce specific pricing and release dates at this time, but the memory and storage shortages have worsened rapidly in the past few months, with limited supply and rising prices affecting the mass production schedule of the three hardware products.

The company stated that it must "re-examine" its shipping arrangements, especially the pricing strategy for Steam Machine and Steam Frame.

Last November, when Valve announced these three hardware products, it revealed to the media that the Steam Machine would be "closer to the entry-level PC market," the Steam Frame would be priced lower than the previously announced $999 Index, and the Steam Controller would be designed to remain competitive in the "advanced input" controller market.

However, within days of its release, RAM prices began to climb rapidly. Valve told the media at the time that pricing was "difficult" because "the market is very strange" and memory prices were "rising in real time."

Since the beginning of this year, PC gamers have generally felt that RAM prices have increased three to four times, mainly because suppliers have devoted more production capacity to the more profitable AI server market.

AMD CEO Lisa Su recently stated in an earnings call that, from a "product perspective," Steam Machine is still progressing as planned, but supply chain pressures are clearly making it difficult to maintain the original launch schedule.

Valve emphasized that the goal remains to launch three products in the first half of this year, but given the rapidly changing supply chain situation, the company needs more time to determine prices and launch dates that it can "confidently announce".

Related reading: Steam Machine's sudden emergence was immediately hampered by memory shortages.

Sony raises full-year profit forecast: Strong content business, hardware costs a concern.

Yesterday, Sony Group announced its third-quarter results for fiscal year 2025 (ending last December). The overall performance was strong, with operating profit and full-year expectations both significantly revised upwards. However, revenue from its film and television business declined year-on-year, showing a structural difference between the coexisting pressures of content business and hardware costs.

• Operating profit for the quarter reached 515 billion yen, a 22% year-on-year increase, exceeding market expectations;
• Sales increased by 1% year-on-year, reaching 3.71 trillion yen.

The company raised its operating profit forecast for this year from 1.43 trillion yen to 1.54 trillion yen, and simultaneously raised its full-year revenue forecast by 300 billion yen to 12.3 trillion yen.

Following the release of the financial report, Sony's Tokyo-listed shares rose by as much as 6%, marking the largest increase since November of last year.

Gaming and online services remain Sony's largest revenue source, but hardware profits are under pressure due to rising prices of core components such as DRAM.

PlayStation's business continued to grow, driven by content, with quarterly software sales reaching 97.2 million copies and PlayStation 5 sales reaching 8 million units. Sony stated that it will offset the impact of rising hardware costs by strengthening software and subscription services.

The music and image sensor business continues to be a growth engine:

Music revenue grew 12.6% year-over-year, benefiting from the recovery of streaming and live performances; image sensor revenue grew approximately 20% year-over-year, primarily driven by demand for high-end smartphones. While a global memory shortage may impact the smartphone industry, Sony believes its high-end sensor products will be less affected.

In contrast, Sony Pictures Entertainment reported quarterly revenue of $2.3 billion, down 12% year-over-year, and operating profit down 11%.

The strong performance of "Venom: The Last Dance" in the same period last year raised the base figure, while the best performing film this quarter was "Chainsaw Man: Reze," which grossed $117 million worldwide. The film division maintains its full-year guidance.

Variety points out that amid market concerns that rising DRAM prices would increase costs and suppress stock prices, Sony's quarterly financial report released a positive signal. The robust performance of the company's content business is offsetting cost pressures on the hardware side.

Sony CEO Hiroki Totoki stated that the company may further adjust its business structure and has announced that it will spin off its television business, which includes Bravia, into a joint venture with TCL Electronics starting next April.

Xiaomi SU7 pushes out latest OTA update: Assisted driving threshold lowered to 300 km/h

Recently, the Xiaomi SU7 received a new OTA update, which focuses on adjustments to the driver assistance functions and adds emergency parking assistance capabilities.

It is reported that this OTA update will upgrade the system version to Xiaomi HyperOS 1.12.0, covering models such as Pro, Max, and Ultra.

The update includes optimizations to end-to-end assisted driving in urban scenarios, and adjustments to lane selection and lateral control logic.

It is worth noting that after the update, the safe mileage threshold for City Navigation Assist has been lowered from 1,000 kilometers to 300 kilometers. Users can unlock the City Navigation Assist function by logging back into their account after accumulating 300 kilometers of Adaptive Cruise Assist, Lane Centering Assist, or Highway Navigation Assist.

Xiaomi Auto stated that lowering the barrier to entry aims to allow users to gradually become familiar with the assisted driving functions, but still emphasizes that "assisted driving is not autonomous driving" and reminds drivers to remain attentive.

Regarding parking assistance, the system has optimized parking planning trajectories, improving the accuracy of parking maneuvers. This update also adds an emergency parking assist function, available for Pro and Max models.

If the driver experiences prolonged fatigue or distraction while using the assisted driving function, and the monitoring system triggers a Level 3 alarm and the driver fails to take over the vehicle in time, the vehicle will automatically activate its hazard lights, sound its horn, and slow down until it comes to a complete stop. At the same time, the assisted driving function will be disabled for 30 minutes.

The vehicle will play a voice prompt when performing an emergency stop to remind the user to take control of the vehicle in time.

GAC Honda responds to reports of the new Fit selling out and production being discontinued: Limited production is not a gimmick; parts supply will continue until 2036.

According to Blue Whale Auto, GAC Honda confirmed yesterday that the new Fit has officially sold out and will no longer be produced, drawing widespread attention to this "legendary car".

In response to questions from the outside world regarding limited production, the scope of production halts, and after-sales service, GAC Honda and relevant informed sources have provided further answers.

The new Fit was launched on January 15th this year with a refreshed price of 66,800 yuan, with a limited run of 3,000 units nationwide.

Sources familiar with the matter revealed that the limited production run stemmed from the model's inherent design and was not a marketing gimmick. With orders nearing their limit on February 3rd, GAC Honda has notified its authorized dealerships nationwide to stop accepting new orders.

The official statement said that the limited edition was intended to reward users who have long supported Fit and provide them with a more exclusive value experience.

Regarding whether the domestically produced Fit will be completely discontinued in the future, GAC Honda stated that as a classic model, the Fit will continue to be developed based on user needs and market trends, but the specific direction will still need to be combined with Honda's global strategy.

Since entering the Chinese market in 2003, the Fit has accumulated sales of nearly 1.5 million units, with the third-generation model contributing the most, with sales of approximately 670,000 units between 2014 and 2020.

With the discontinuation of the new Fit, many consumers are worried about after-sales service.

GAC Honda responded that, according to the "Automobile Sales Management Measures", parts for discontinued models must be supplied for at least 10 years, therefore, parts for the new Fit will be supplied until 2036.

The official delivery schedule was also disclosed: production and delivery of 3,000 units will be completed from January to April, with approximately 600 units expected to be delivered in January and February, approximately 800 units in March, and all deliveries completed in April.

Vidu, a digital technology company, has completed a Series A+ financing round of over 600 million yuan.

Yesterday, Shengshu Technology announced the completion of its Series A+ financing round, raising over 600 million RMB.

According to official information, this round of financing was led by Zhongguancun Science City Company and Xinglian Capital, with strategic investments from listed companies such as Wondershare Technology, Visual China, and TRS, and existing shareholders such as Qiming Venture Partners and Beijing Artificial Intelligence Industry Investment Fund continuing to increase their investment.

The core team at Shengshu Technology comes from the Institute for Artificial Intelligence at Tsinghua University, and is one of the world's earliest teams to research multimodal generation algorithms. The company has continuously iterated its expertise in text-based and image-based video, launching Vidu Q1, Q2, and Q3 successively.

In the latest rankings by the internationally authoritative organization Artificial Analysis, Vidu Q3 ranked first in China and second globally, tied with xAI Grok, and ahead of Runway Gen-4.5, Google Veo 3.1, and OpenAI Sora 2.

In terms of commercialization, Vidu Technology has built a product matrix including Vidu MaaS, Vidu SaaS, Vidu Agent, and Vidu APP, covering more than 200 countries and regions around the world.

Counterpoint: Global average selling price of smartphones surpasses $400 for the first time in a single quarter.

Yesterday, market research firm Counterpoint Research released its latest research report, showing that the global smartphone market saw significant growth in the fourth quarter of last year, with the average selling price (ASP) exceeding $400 (approximately RMB 2,776) for the first time, and revenue increasing by 13% year-on-year to a record high of $143 billion.

The research report points out that although shipments only increased by 5% year-on-year, the trend towards premiumization, holiday demand, and promotional activities jointly drove the overall value increase.

Apple continued to solidify its market leadership in the quarter, with shipments and revenue increasing by 14% and 23% year-over-year, respectively, both reaching record highs.

The iPhone 17 series became the main driver of growth, with the Pro version showing a significant increase in market share, bringing Apple's global revenue market share to 59%.

Jeff Fieldhack, research director at Counterpoint, pointed out that Apple's revenue hit record highs in the United States, China, Latin America, Western Europe, the Middle East, and South Asia.

Samsung, with a year-on-year growth rate of 17%, became the fastest-growing manufacturer in terms of shipments among the top five brands, with revenue increasing by 12% year-on-year.

The Galaxy A series performed strongly in several emerging markets, and the Galaxy S25 and Z Fold7 also received positive feedback in some regions. However, due to the increased market share of the A series, Samsung's ASP (Average Selling Price) decreased by 4% year-on-year.

As for other major manufacturers, Xiaomi's revenue and shipments decreased by 9% and 11% year-on-year, respectively. Affected by rising component costs and supply constraints, its performance in the Chinese and Indian markets was under pressure.

OPPO's revenue grew by 23% year-on-year, driven by the expansion of its high-end product portfolio and its expansion into emerging markets, while its average selling price (ASP) increased by 6%. Vivo, on the other hand, achieved a 6% year-on-year revenue growth, relying on its solid performance in the Chinese and Indian markets.

Counterpoint Research Director Tarun Pathak predicts that smartphone ASPs will continue to rise this year, driven by premiumization, demand for AI features, and rising memory prices. However, high component costs may squeeze shipment volumes, forcing manufacturers to focus more on value growth and product mix optimization.

The Milan Winter Olympics created its first official Olympic large-scale model based on Alibaba's Qianwen platform.

Yesterday, the International Olympic Committee announced the creation of the first official large-scale model in Olympic history based on Ali Qianwen, which has already been put into use in the Games administration and public services system.

The International Olympic Committee stated that this Winter Olympics will be the "smartest" ever, with AI technology integrated into event organization, content production, and spectator services.

The International Olympic Committee has launched the "National Olympic Committee AI Assistant" for national delegations. Leveraging the multilingual capabilities of the Thousand Questions feature and the official manual's corpus, it provides delegations with native-language answers to questions regarding eligibility verification and logistical scheduling, thereby improving the efficiency of cross-border collaboration.

The "Olympic AI Assistant," accessible to the public, will provide real-time Q&A services on Olympics.com, including event rules and historical data, further bridging the gap between viewers and the events.

In terms of content production, the Qwen-VL-based automatic media description system has been put into live broadcast production, which can automatically identify key events such as goals and fouls and generate descriptions; the Milan Winter Olympics Organizing Committee also generated promotional materials in batches based on Alibaba's Wanxiang big data model, enabling AIGC to be applied on a large scale for the first time in Winter Olympics content production.

Alibaba Cloud has deployed AI-enhanced special effects technology in 10 core venues, using multi-model fusion algorithms to solve the problem of snow scene reconstruction, bringing clearer "bullet time" and "time slice" images.

In addition, Alibaba Cloud also supports the event's traffic management system and energy tracking system, improving access in mountainous areas and enhancing the event's sustainable operation. International Olympic Committee President Coventry stated that AI-driven intelligent capabilities will be a "lasting gift" from the Milan Winter Olympics to the world.

Starting at 17,000 yuan, Jia Yueting unveils his first batch of embody intelligent robots.

Yesterday, Faraday Future (FF) unveiled its first batch of Embodied Intelligence (EAI) robots at the National Automobile Dealers Convention (NADA) in Las Vegas, USA, and simultaneously opened sales and pre-orders.

At the press conference, FF launched three major series of robots:

• The Futurist series is positioned as a full-size professional humanoid robot, with prices starting at $34,990 (approximately RMB 242,000);
• The Master series consists of athletic humanoid robots, priced from $19,990 (approximately 138,000 yuan).
• The Aegis series are security and companion quadruped robots, priced from $2,499 (approximately 17,000 yuan).

Among them, the Aegis series can be equipped with a four-wheel version based on the four-leg structure, and the boom series products are expected to be released in the second quarter of this year.

This announcement sparked considerable discussion. According to reports from 21st Century Business Herald and Beijing Business Today, netizens generally questioned whether FF's car manufacturing business was yet on track, and whether its robotics business possessed genuine technological accumulation and mass production capabilities remained to be seen.

Some comments bluntly stated that the product's appearance "clearly resembles a knock-off" or "looks like something from Huaqiangbei (a major electronics market in Shenzhen)."

21st Century Business Herald pointed out that FF has faced a long-term operational crisis since its listing. In April last year, Jia Yueting returned to the position of co-CEO in an attempt to promote strategic restructuring.

During CES 2026, FF announced its five-year business plan, outlining mass production targets for models such as the FX Super One, and emphasizing that it will build its future growth curve through a "dual-track" approach of automobiles and robotics.

Jia Yueting himself also posted on Weibo that the three major product series and their "killer price" will drive the EAI robot industry to accelerate its transformation towards "four modernizations" and that FF will start to grow faster from this year.

However, since he went to the United States to start his business in 2017, he has not returned to China for eight years and is still burdened with huge debts. The timing of his return to China remains a focus of public opinion.

Related reading: Despite being in the US for eight years, Jia Yueting still managed to release three robots before the Spring Festival Gala.

After more than 30 years of maintenance, the developers of sudo are calling for funding.

According to reports from The Register and Linuxiac, sudo, a key open-source tool that has been maintained for over 30 years, is currently facing resource shortages, and its long-time maintainer, Todd C. Miller, is publicly seeking institutional or individual sponsorship to sustain the project's continued development.

Miller took over the maintenance of sudo in 1993.

sudo is one of the most fundamental permission management tools in Unix and Linux systems. It allows users to execute commands under controlled policies as other users and is a core component of the security models for servers, desktops, and embedded systems.

Miller stated that since his former employer, Quest Software, ended its sponsorship in February of last year, he has been relying on his own efforts to maintain updates, but the lack of stable funding has significantly slowed down the development pace.

Despite this, sudo updates have not stopped. According to the project update log, Miller is still continuously fixing vulnerabilities and cleaning up code.

Over the past few years, sudo has been exposed to several high-risk security issues, such as the heap buffer overflow vulnerability discovered in 2021, which allowed local users to obtain root privileges without authorization, highlighting the importance of continuous maintenance.

Meanwhile, the Rust rewrite sudo-rs is gaining wider adoption. Ubuntu made sudo-rs the default implementation in Ubuntu 25.10, released last October.

Miller stated that he trusts the sudo-rs team and believes that sudo-rs may become mainstream in the next few years, but he still hopes to continue maintaining the original sudo and emphasized that he will not easily hand over the project to strangers, especially after the xz backdoor incident.

Miller emphasized that he has no intention of abandoning sudo, but due to a lack of funding, maintenance work has been forced to focus on fixing problems rather than adding new features.

He pointed out that open-source maintainers generally face the dilemma of exhaustion of energy and resources, and the current state of sudo is a microcosm of this structural problem.

Li Xiang: The ultimate form of a car is a robot.

Yesterday, Li Xiang, CEO of Li Auto, posted on Weibo, systematically reviewing the technological path of the past decade with "embodied intelligence" as the core narrative, and taking the new Li Auto L9 as the key starting point for the next stage.

Li Xiang stated on Weibo that when Li Auto was founded in 2015, it regarded "robots" as the ultimate form of automobiles , rather than simply faster and smarter means of transportation.

He outlined the progress of Li Auto in areas such as perception, operating system, central domain controller, self-developed chips, and self-developed large models in a timeline format.

In response to external doubts about whether "going all in on AI means neglecting the car," Li Xiang emphasized that embodied intelligence must be integrated into a good car , and stated that he will still devote 70% of his time to the research and development of the car itself.

Li Xiang stated that the all-new Li Auto L9 will evolve from a "car" into an "intelligent agent," possessing a complete technology stack of "eyes, brain, heart, nerves, hands and feet." It is capable of recognizing users, understanding users, and proactively serving them, and is regarded as a key product for Li Auto to move towards becoming an embodied intelligent robot.

The interior of the Zhijie V9 has been revealed, featuring a debut color scheme of Velvet Purple and rotating seats in the second row.

Yesterday, HarmonyOS unveiled details of the cabin and interior of its first flagship MPV, the "Zhijie V9," highlighting the debut of the "Velvet Purple" color scheme, the new second-row rotating seats, and a number of high-end comfort features.

The official video shows that the new car's interior uses materials such as natural wood, NAPPA top-grain leather, and crystal-cut crystal, and the front row is equipped with an integrated screen.

The second-row seats can rotate 45°, 90°, and 180°, and can achieve a 123° Eames reclining mode. They are also the first to be equipped with wraparound airbags, further enhancing occupant protection.

According to Autohome, in terms of power, the Zhijie V9 is equipped with Huawei's Snow Owl intelligent range extender system, with a maximum combined range of over 1250km (fully fueled and fully charged, 53kWh battery version).

According to information submitted to the Ministry of Industry and Information Technology, the extended-range electric vehicles will offer two battery packs: 37 kWh lithium iron phosphate and 53 kWh ternary lithium. The WLTC pure electric range will cover 158 km to 223 km, with a minimum fuel consumption of 6.8L/100km when the battery is depleted. Some higher-spec versions will be equipped with dual motors, but specific motor parameters have not yet been released.

The vehicle measures 5359×2009×1879mm with a wheelbase of 3250mm, and the official description states that its interior space "far exceeds that of its competitors." The new car also features a welcome light carpet, an in-car oxygen bar, and is equipped with Huawei's next-generation intelligent driving assistance hardware.

Canon releases PowerShot G7 X Mark III 30th Anniversary Limited Edition for $1299

According to The Verge, Canon announced yesterday the launch of the PowerShot G7 X Mark III 30th Anniversary Limited Edition camera to commemorate the 30th anniversary of the PowerShot series since its debut in 1996.

The new product is priced at $1,299 (approximately RMB 9,015) and will be available for purchase in April.

This limited edition maintains the same core specifications as the currently available PowerShot G7 X Mark III, with the main changes focused on the appearance and accessories. The camera features a graphite gray finish, a knurled control ring, and a 30th-anniversary logo, and comes bundled with a Peak Design limited-edition wristband and a 32GB SD card.

The PowerShot G7 X Mark III was originally released nearly 7 years ago, targeting the video blogger market. It supports vertical video shooting, features a 20.1-megapixel sensor, a 3-inch flip-out touchscreen, an equivalent 24 to 100mm f/1.8 to 2.8 lens, and provides a dedicated microphone jack.

Video capabilities include uncropped 4K 30fps and 1080p 120fps slow-motion shooting, and support for live streaming directly to YouTube via Wi-Fi.

This model previously saw a surge in resale value due to supply shortages and its popularity among online influencers. After Canon increased production last August, the official price was adjusted to $880. This limited edition, however, is priced even higher, emphasizing its collectability.

Spotify enters physical book retail with cross-media "page number matching" feature.

According to TechCrunch, Spotify announced that it will launch a physical book purchase feature in the US and UK this spring, allowing users to buy print versions of audiobooks within the app. This marks the first time the platform, which has long focused on digital content, has entered the physical book retail sector.

Spotify has partnered with Bookshop.org, which handles pricing, inventory, and delivery. Users can click "Add to Shelf" within the app and be redirected to the Bookshop website to complete the purchase. Each order generates revenue for the local independent bookstore.

At the same time, Spotify launched several new features around the audiobook experience to strengthen its role in the reading ecosystem:

• "Page Match" allows users to scan any page of a print book or e-book, and the system will automatically jump to the corresponding position in the audiobook, achieving a seamless switch between listening and reading.
• "Scan to Listen" and "Scan to Read" allow users to switch between audio and physical books. These two features are currently available for English books and are expected to be rolled out to all audiobook users by the end of February.
• "Audiobook Recaps," previously only available on iOS, will be coming to Android this spring. This feature generates a short recap based on where the user last stopped reading, helping them quickly resume their reading pace.

On the business front, Spotify's audiobook segment continues to grow.

Since its launch two years ago, the platform has seen a 36% year-on-year increase in audiobook users and a 37% increase in listening time. More than half of its 281 million paid subscribers have listened to at least one audiobook. The platform currently offers over 500,000 audiobook titles. Spotify is expected to release its Q4 2022 earnings report on February 10th.

The report suggests that this expansion of physical book sales and the launch of cross-media features indicate that Spotify is attempting to move from a single digital audio platform to a complete reading ecosystem covering print books, ebooks, and audiobooks, and to seek new growth opportunities beyond traditional booksellers such as Amazon and Barnes & Noble.

Relevant departments: Rectifying the abuse of platform rules such as "refund only" and "mandatory shipping insurance".

According to Xinhua News Agency, yesterday, the State Administration for Market Regulation introduced the latest progress in the governance of the platform economy at a press conference held by the State Council Information Office, and responded to concerns about the abuse of platform rules such as "refund only", "mandatory freight insurance" and "lowest price on the entire network".

Zhu Jianqiao, Director of the Department of Online Transaction Supervision and Management of the State Administration for Market Regulation, said that while the platform economy is expanding rapidly, it has also exposed prominent contradictions such as abuse of rules, lack of transparency in fees, and false marketing.

Last year, regulatory authorities launched a concentrated campaign to address issues such as platform rules, fees, irregularities in live-streaming e-commerce, and disorderly competition in the food delivery industry, achieving initial success.

Regarding platform rules, regulatory authorities have rectified unreasonable rules such as "refund only," "mandatory shipping insurance," and "lowest price across the entire network" through methods such as interviews and rectification, and have strengthened the protection of the rights and interests of small and medium-sized businesses and consumers.

Meanwhile, regulatory authorities, together with the Cyberspace Administration of China, issued the "Measures for the Supervision and Management of Rules for Online Trading Platforms" and introduced guidelines for fee compliance and platform compliance management to promote the formation of a long-term regulatory mechanism.

In the live-streaming e-commerce sector, regulatory authorities have launched special campaigns to address issues such as counterfeit and substandard products and false advertising. They have investigated and dealt with illegal cases involving companies such as Chengdu Kuaigou Technology Co., Ltd., and provided guidance to local authorities on handling cases of false advertising by top live-streamers such as "Taiyuan Laoge" and "Da LOGO". A total of 30 typical cases have been released in four batches, and the "Supervision and Management Measures for Live-streaming E-commerce" have been jointly issued with the Cyberspace Administration of China.

In the food delivery industry, in response to irrational competitive behaviors such as "verbal battles" and "subsidy wars" that occurred last year, regulatory authorities conducted administrative interviews with major platforms, clarified regulatory red lines, and issued the national standard "Basic Requirements for Service Management of Food Delivery Platforms" to promote standardized promotions and rational competition in the industry.

Zhu Jianqiao stated that the next step will focus on strengthening the main responsibility of platforms, improving systems and mechanisms, enhancing law enforcement constraints, and promoting intelligent supervision. A number of typical cases will be investigated and exposed through methods such as listing cases for supervision, upgrading jurisdiction, and designating jurisdiction, to further maintain fair competition and promote a greater role for the platform economy in high-quality development.

Meituan acquires Dingdong Maicai's China business for $717 million

According to Jiemian News, Meituan announced on the Hong Kong Stock Exchange yesterday that it will acquire 100% equity of Dingdong Maicai's China business for an initial consideration of approximately US$717 million. After the transaction is completed, Dingdong will become an indirect wholly-owned subsidiary of Meituan, and its financial results will be consolidated into the Meituan Group.

The announcement indicates that both parties have signed a share transfer agreement outside of trading hours, and the initial consideration is subject to adjustment. The transferor may extract up to US$280 million from the target group, provided that the target group's net cash at the time of closing is not less than US$150 million. Dingdong Maicai's overseas business is not included in this transaction and will be divested before the closing.

Dingdong Maicai, founded by Liang Changlin, is a leading fresh food instant retail platform in mainland China, with its main feature being "delivery in as little as 29 minutes".

As of September last year, Dingdong operated over 1,000 pre-positioned warehouses in China, with over 7 million monthly active users, boasting a high repurchase rate and direct supply chain sourcing capabilities. In the third quarter of 2025, Dingdong achieved a record high revenue of RMB 6.66 billion and achieved GAAP profitability for seven consecutive quarters.

In its announcement, Meituan emphasized that Dingdong aligns closely with Meituan's long-term strategy in terms of product strength, supply chain, technology, and operations, further strengthening Meituan's presence in the on-demand retail sector. Meituan has been continuously developing retail formats such as Elephant Supermarket in recent years to meet the growing demand for on-demand retail from Chinese consumers.

The 2026 CCTV Spring Festival Gala will partner with Bilibili to enable real-time bullet comments.

Yesterday, the 2026 Spring Festival Gala announced a partnership with Bilibili (B-site), making Bilibili the exclusive danmaku (bullet screen) video platform for this year's Year of the Horse Spring Festival Gala.

This marks the second collaboration between the two parties. Bilibili has also upgraded its products and technology for the Spring Festival Gala live stream to enhance the interactive experience of bullet comments. According to official sources, viewers can participate in real-time discussions about the program through Bilibili's bullet comments during this year's Spring Festival Gala live stream.

Meanwhile, starting before the Spring Festival, users can also create interest-based rooms on Bilibili to watch Spring Festival-related content with friends and discuss the Spring Festival Gala programs. Bilibili expects the number of interest-based rooms to exceed 1,000 during this year's Spring Festival.

Last year, the Spring Festival Gala was watched by over 100 million people on Bilibili, with 83% of them being under 30 years old. The co-created program "Spring Red Envelopes" garnered over 1 billion views across the internet, becoming the second most-watched Spring Festival Gala program overseas.

After the replay of the 2025 Spring Festival Gala was uploaded to Bilibili, it garnered over 300 million views, and related fan-made content accumulated over 2.54 billion views, giving rise to numerous works with tens of millions of views.

OPPO's New Year film "The Time Thief" is now available online: Filmed entirely with Find X9 Pro

Yesterday, OPPO's New Year film "The Time Thief" was officially released online. The entire film was shot using an OPPO Find X9 Pro.

The film, co-created by director Yang Lina of "Little Me" and screenwriter Qiu Yujie of "The Beginning", focuses on the emotional dilemmas between young people and their families today. Using "stealing time" as a metaphor, it tells a spiritual journey that causes chaos in a small town due to an accident, but ultimately leads to self-awakening and a return to the present moment.

"Supreme Matti" Confirmed for Import

According to Mtime, "Marty the Great" has been officially confirmed for release in mainland China.

The film tells the story of shoe store clerk Marty Mauser's desperate pursuit of his dream to win the world table tennis championship, and is known for its strong visual style and highly dramatic narrative.

In the film, Marty is willing to set aside his dignity and travel far and wide to prove himself, all for a chance to return to the track. Whether he can use his obsessive passion to shatter the doubters and reach the top becomes the core suspense of the story.

"Marty the Great" performed brilliantly in North America, topping the box office charts for A24 Pictures films upon its opening, and garnering nine nominations at the 98th Academy Awards, including Best Picture, Best Actor, Best Director, and Best Editing.

#Welcome to follow iFanr's official WeChat account: iFanr (WeChat ID: ifanr), where more exciting content will be presented to you as soon as possible.

ifanr | Original Link · View Comments · Sina Weibo

The GPT-5.3-Codex released early this morning can be seen as a powerful counterattack by OpenAI against the recent surge in popularity of various local agents, and primarily a counterattack against Anthropic.

With the Codex desktop application released by OpenAI a few days ago, the functions that popular tools like Skill, Cowork, Claude Code, and even Openclaw can now be achieved through the Codex shell and the capabilities of the GPT-5.3-Codex model.

Similar to how we introduced Cowork's capabilities before, we also assigned Codex some similar tasks to complete, such as directly processing local files, converting various formats, combining different skills, creating Word/PPT/Excel files, downloading videos, developing apps, etc.

GPT-5.3-Codex performs remarkably well. Compared to installing Claude Code from scratch, downloading Codex directly is a better option for new users. This is also a trend among model vendors in the future; initially, everyone started with the cumbersome command-line terminal to create local agents, but they are gradually returning to user-friendly visual interfaces.

Online reviews of Codex have reversed considerably in the past few days, with many developers switching from Claude Code to Codex. Some independent developers in China have also stated that Codex Plus members can use it without the ruthless account bans that Claude often does.

Ultraman excitedly announced that Codex had surpassed 1 million active users. His model update blog was also filled with unreserved and generous praise.

GPT-5.3-Codex is our first self-built model . By using 5.3-Codex, we were able to release it at such a rapid pace.

Working with Claude's team for two weeks, using Claude Code—100% AI code—to create something like Cowork; and then there's OpenAI's article from the end of last year, "Building an Android version of Sora in 28 days using Codex"—the era of agents has truly arrived.

Replace my ChatGPT and Claude Code with Codex

Like most local agents, whether in the terminal or Cowork, we first select a working folder. In Codex, we can create multiple projects, select the corresponding folders, and then begin the conversation; Codex calls these Threads.

Let's start with the most common and simple example. We added an empty download folder, then clicked "Start a thread" and selected the GPT-5.3-Codex model; just like having a conversation in ChatGPT, we entered the commands.

We asked Codex to download an X video for us. Codex automatically checked for available skills to handle it and then downloaded it using the yt-dlp tool. The video was over four hours long, and Codex kept updating the download progress in the dialog box.

After downloading the video, we can also ask it to extract the transcript of the video, provide us with a bilingual document, and finally package the entire process into a Skill for easy use next time.

If you have interesting clips in a video and want to trim the video or convert the trimmed video into a GIF, you can do it all in Codex.

For example, we download a video and then ask it to crop 5s-25s of the video to create a new video. Thanks to GPT-5.3-Codex's fast token processing, the whole process doesn't take long; it depends more on the local computer's hardware decoding and encoding capabilities.

Alternatively, we can directly ask it to convert the first 5 seconds of the video into a GIF file, ensuring the size is within 10MB, the frame rate can be adjusted, and the width is controlled at 640px for clarity.

Soon, we'll get the corresponding GIF file. Even more extreme, it can convert the entire video into images at 30 frames per second, with each frame being an image.

These direct processing capabilities for local files, along with GPT-5.3-Codex's excellent performance on the Terminal-Bench-2 test suite, enable Codex to essentially fulfill the functional requirements of various productivity and efficiency tools.

In comparison, the newly released Claude Opus 4.6 scored 65.4% on Terminal-Bench 2.0 and 77.3% on GPT-5.3-Codex.

For example, if there are multiple images in this folder, our first requirement is to rename these image files according to their content, and keep the filenames to no more than 20 characters, without allowing the use of symbols.

After the automatic modification is complete, we can also ask it to stitch these images together, whether vertically or horizontally. Codex can do this by calling the corresponding tools.

Like Claude Skills, Codex can install a wide range of skills from the Skills Marketplace, and it already offers several skills within the application, including pptx, xls, word, canvas, and notion.

Returning to basic programming skills, the upgraded GPT-5.3-Codex performs significantly better than GPT-5.2. We directly asked it to write a "Word of the Day" app. Unlike ChatGPT, which simply provides a non-transferable webpage using Canvas, Codex allows you to build the project locally from scratch and then deploy it to a webpage using skills like Vercel or Cloudflare.

Here we selected the Extra High reasoning mode. Before each operation, GPT-5.3-Codex would ask me for the next operation option. This is also related to the fact that Codex can directly call different skills according to the task situation. Among them, the Brainstorming Skill will automatically conduct a continuous dialogue mode.

Ultimately, it essentially fulfilled all the features I initially asked it to have, and it can also be further developed for macOS, iOS, and Android versions.

If we have an existing code project, we can also select the project folder and open it in Codex. GPT-5.3-Codex will analyze the bugs in the project and fix them.

For a long time, Anthropic's Sonnet/Opus model and Claude Code tool were the preferred choices for developers, both in terms of tools and models. OpenAI's lagging behind in programming, especially in long-code logical reasoning, led many developers to switch sides.

The emergence of GPT-5.3-Codex was intended to end this debate. Now, GPT-5.3-Codex not only crushes its own predecessor in programming benchmarks and real-world performance, but also shows signs of thoroughly outperforming competitors' models. It truly possesses the ability to write, test, and infer code.

The game project was a key case study in the website development section of this model introduction blog. We also had GPT-5.3-Codex create a simple physics-based pinball game. While the overall result didn't meet my expectations—I specifically requested an RPG in the hints—the interface provided by GPT-5.3-Codex was still too rudimentary. However, it's still playable.

We also found some small games on X made with GPT-5.3-Codex, like this Super Mario-like coin-collecting game.

There are always stronger players among the strong.

For Anthropic, what OpenAI is doing today is something we've already done. Whether it's code, agent capabilities, or starting to develop local agents, transitioning from the previous Codex terminal to macOS apps.

In the realm of technology, OpenAI seems to be following in Claude's footsteps. Claude has been deeply involved in coding, while OpenAI has been working on Sora, daily reports, browsers, and ChatGPT agents, but none of them have made much of a splash, so they have also started to focus on coding. Claude launched Cowork in early January, and OpenAI followed suit in early February by releasing the Codex App.

Just like today's flurry of releases, at 1:45 AM, Claude officially launched Claude Opus 4.6, followed by OpenAI's release of GPT-5.3-Codex. Both models are actually designed to give agents more powerful foundational capabilities. Previously, it was about code/vibe coding, but now, for an agent to perform well, it's basically about "writing good code".

While Opus 4.6 performs even worse than Opus 4.5 on SWE-Bench and is not as strong as GPT-5.3-Codex on Terminal-Bench 2.0, Opus has unprecedentedly increased the context length to a window of one million tokens. Moreover, the performance of these benchmarks is not significantly different.

Claude said, "My Sonnet 5 hasn't been released yet; that's where the real skill comes in."

We also found some of the latest test cases for Opus 4.6 online. Some netizens said that Claude 4.6 Opus completely refactored his entire codebase with just one call, modularizing the original messy "mountain of crap" of code. No other model can do this as well as Opus.

Some netizens compared Opus 4.6 and 4.5, having both models play the same management game to see whose account level, wealth, and equipment were higher. The test blogger mentioned that version 4.6 took longer to formulate strategies in the early stages, but made better strategic decisions and ultimately achieved a significant lead.

Another netizen also created a game, but it's a clone of Pokémon. The blogger mentioned that this is the coolest thing he's ever made using AI. He mentioned that Claude Opus 4.6 took 1 hour and 30 minutes to develop, used 110,000 tokens, and only iterated three times.

In the official CLaude demo and early user feedback, a case of Opus performing exceptionally well was mentioned. Opus 4.6 autonomously closed 13 issues (unresolved issues in the project) within a single day and accurately assigned another 12 issues to the correct human team members.

Like the agent swarm in Kimi K2.5, Opus 4.6 can also manage the codebase of an organization of 50 people. In Claude Code, we can form Agent Teams, summoning an entire team of AI, instead of a single AI fighting alone. These AIs can be responsible for writing code, reviewing, and testing, and they can collaborate autonomously.

Some users also tested the Agent swarm in Claude Code and mentioned that after enabling the swarm, Opus 4.6 was 2.5 times faster and performed better.

Our current situation is just like this picture; although one mountain is higher than another, we can't escape this cycle. Gemini probably stole the spotlight a few months ago, Claude should be next in January, and then it looks like it'll be OpenAI's turn, or Musk's Grok.

Fortunately, during this cycle, we, as users, can clearly feel that the capabilities of AI are constantly improving.

The API for GPT-5.3-Codex is not yet open because the model is too powerful and poses a significant risk. Therefore, OpenAI is still considering how to securely enable the API.

Claude Opus 4.6 can be used in multiple ways, including the Claude universal chat application, Claude Code, and API. As two of the first models released by the three major international vendors this year, they are definitely worth trying.

In the future, better serving agents and enabling them to do things for us will remain a key focus of major model updates.

#Welcome to follow iFanr's official WeChat account: iFanr (WeChat ID: ifanr), where more exciting content will be presented to you as soon as possible.

ifanr | Original Link · View Comments · Sina Weibo

In 2016, DJI released the first-generation Mavic Pro, while smartphone manufacturers were engaged in a fierce battle over the pixel count of dual cameras.

No one could have imagined that ten years later, these two seemingly parallel lines would collide head-on at the point of "handheld gimbal cameras"—

vivo has confirmed that it internally initiated a Vlog camera project at the end of 2025, aiming to compete with DJI's Pocket series, and it is expected to be released in 2026.
Coincidentally, according to iFanr, InStone's handheld gimbal camera will also be released in the first half of the year.

According to conventional thinking, cameras and mobile phones are clearly distinct and completely different industries. However, if you focus on the Pocket form factor, you will find that the entry of mobile phone manufacturers such as vivo and InSight is not abrupt, but rather timely and logical.

Seemingly a crossover, but actually from the same field.

vivo's confidence in entering the camera market stems primarily from its supply chain.

After two generations of refinement with Pocket 1 and 2, Pocket 3 uses a one-inch CMOS sensor. Although DJI has not explicitly stated which company this sensor comes from, industry clues almost universally point to Sony.

If you are interested in mobile imaging, you should have already keenly thought of a few familiar models – the Sony IMX989 sensor widely used by domestic mobile phone manufacturers and its subsequent iteration, the LYT900 CMOS sensor.

Looking back over the past few years, how much real money have domestic mobile phone manufacturers poured into Sony's one-inch sensor? Hundreds of millions of yuan in R&D expenses have been used to fine-tune color science, optimize the dynamic range of HDR, and squeeze noise control in extreme night scenes.

In some ways, mobile phone manufacturers understand this sensor better than camera manufacturers.

With this sensor at its core, essentially you remove the phone's communication baseband, replace the power-hungry OLED screen with a small viewfinder, and relax the heat dissipation restrictions—congratulations, you've got a vlog camera with overkill performance.

vivo even has its own "brain" ready-made. Its self-developed Blueprint V3+ image chip has the ability to process portrait videos at 4K 60fps, and can independently handle image processing tasks.

In addition to the foundation provided by the supply chain, there are also some achievements left by previous bold explorations, which have become an excellent foreshadowing for vivo's entry into the handheld gimbal camera market.

For handheld gimbal cameras, the soul lies in the "gimbal," and image stabilization is paramount. vivo attempted to incorporate a miniature mechanical gimbal into the phone during the X50/X60 era. Although they didn't continue with this approach, this experience provided them with crucial data on miniature mechanical structure design and drop resistance testing.

Beyond the gimbal, there's also the focal length. Vlogs rely on ultra-wide-angle lenses to capture selfies, and DJI's Pocket evolved from 28mm in the first generation to 20mm in the second to address this issue. In the mobile phone industry, vivo is one of the very few manufacturers that has shifted to using ultra-wide-angle lenses for primary video tasks—the vivo X200 Ultra sets its main camera at a 35mm focal length, more suited to street photography, while the 13mm ultra-wide-angle lens becomes the default main video camera, allowing vivo to accumulate considerable experience in ultra-wide-angle video.

Hardware determines the lower limit, while post-processing determines the upper limit. DJI Pocket's popularity stems from D-Log M, and vivo was also one of the first manufacturers to integrate Log curves and support the ACES industrial-grade color encoding system, meaning its footage is capable of being used in professional software post-processing.

By piecing together the micro-gimbal structure, ultra-wide-angle algorithm, and Log workflow, vivo was able to assemble a mature Pocket camera with almost no need to reinvent the wheel.

We have found on the State Intellectual Property Office's public website that vivo has applied for patents related to this type of device:

From the perspective of mobile phone users, a Pocket device launched by mobile phone manufacturers can also solve many problems.

As a personal terminal, mobile phones are not solely for video recording; they also have to handle many complex tasks in life and work. If you want to film a complete performance, you need to worry about the phone's battery level, and when you want to create, you may be disturbed by WeChat pop-ups.

In addition, the ever-expanding WeChat chat history, often reaching tens of gigabytes, makes mobile phone storage space scarce, especially with the soaring price of memory today. Large memory phones are also becoming more and more expensive, leaving little space for images.

Handheld gimbal cameras are not sensitive to onboard memory—inserting an SD card that costs a few tens of yuan can handle hundreds of gigabytes of footage.

This is a perfect "image-specialized" container. Its algorithms and UI are highly consistent with those of mobile phones, while separating the "recording" function from the phone. This not only alleviates the anxiety about the phone's battery life and storage, but also provides a purer creative environment.

With both hardware and software in place and a complete cultural foundation, there is an even grander keyword, which is the killer feature of mobile phone manufacturers and the moat that DJI has yet to build: ecosystem.

Imagine if cameras were no longer isolated entities, could footage be transferred to a mobile phone with a single tap using a proprietary protocol?
Is it possible to directly access those AI models on your phone that have been trained millions of times for one-click editing?
Even during filming, can the phone and camera work together, with one serving as the main camera and the other as the close-up camera?

The ability to seamlessly integrate a standalone camera into the entire system ecosystem is the most lethal "killer app" in the hands of mobile phone manufacturers.

The Normandy moment in Pocket form

We used vivo as an example and spent thousands of words arguing for the rationale behind vivo's entry into the handheld gimbal camera market. However, if we shift our focus to the entire domestic mobile phone industry, we will find that many companies actually have similar conditions to some extent.

In other words, vivo is by no means the only suitable entrant; Huawei, Xiaomi, OPPO, and Honor all have similar conditions.

Indeed, this is the case. Multiple sources have reported that Huawei and Xiaomi are also continuously evaluating this direction, while OPPO was reported as early as the end of 2025 to have its Chief Product Officer and Senior Vice President Liu Zuohu personally leading the Pocket project, with the actual product expected to be released in 2026.

Even more radical was Honor's approach. A month ago, at CES 2026, the world's largest consumer electronics show, which opened in Las Vegas, Honor directly showcased its even more aggressive Pocket solution—the Robot Phone. This solution integrates the gimbal and camera directly into the phone, folding them into the imaging module. For more details, please click on our report: https://mp.weixin.qq.com/s/NCiPnXbK2L-zqMlV3VrXKw

The collective shift away from their core businesses is due not only to favorable objective conditions, but also to the urgent need for new growth drivers among mobile phone manufacturers.

Looking back over the past few years, after experiencing the work-from-home boom in 2020, the tablet market quickly entered a long period of correction. According to Qmdia data, global tablet shipments fell to their lowest level in nearly five years in 2023. Although there was a recovery in 2024 and 2025, the overall market has already reached its ceiling.

Smartphone manufacturers need new narratives, and DJI's Osmo Pocket 3 has proven the market's hunger with its sales figures.

As of September 2025, total sales of the Osmo Pocket 3 exceeded ten million units, achieving this astonishing figure in just two years.

This number may seem insignificant in the face of the mobile phone market, which often reaches hundreds of millions, but if we break it down by model, it is already on par with the annual sales of most manufacturers' core flagship products, proving that handheld gimbal cameras are definitely a market with potential.

Considering its unit price, high profit margin, and extremely high user stickiness, this is definitely a piece of braised pork that's incredibly lucrative.

For Huawei, Xiaomi, OPPO, and Vivo, the objective conditions have long been ripe. DJI has also proven the value of this blue ocean market with its sales of tens of millions of units. If they don't participate, it would be tantamount to handing over their most core and highest-quality imaging users to others.

In the highly competitive ecosystem of the mobile phone industry, the balance is often fragile. As soon as one manufacturer enters the market and successfully establishes a path, this "balance" will be instantly shattered. The remaining players cannot stand idly by and watch their competitors monopolize an ecosystem niche that can be highly integrated with mobile phones. It can be said that this is an inevitable arms race.

Besides mobile phone manufacturers, there is actually another strong competitor in the Pocket market – Inspiron.

Previously, there were reports that Inspire had designed a modular Pocket product, but it has yet to be launched. However, as of the time of this article's publication, several insiders have confirmed to Jiemian News that Inspire's handheld gimbal camera will be released in the first half of the year and will be available in multiple colors. It will be positioned to compete with DJI's Pocket and may adopt a modular design and dual cameras. The industry may soon see a handheld gimbal camera with a mid-to-long telephoto lens.

Finally, I'd like to talk about the Pocket format itself.

In fact, this is an underrated form factor innovation. Before its emergence, mobile phone image stabilization relied heavily on cropping and the use of stabilizers, which was not portable enough. Moreover, as a personal terminal, mobile phones needed to undertake a lot of other tasks and could not focus on shooting. Although camera image quality was good, it was even worse in terms of portability and ease of use, and the cost was high, making it a high barrier to entry.

The genius of the Pocket lies in its use of a mechanical gimbal to solve the image stabilization problem, its extremely small size to solve the portability problem, and its ability to minimize friction during shooting. There is no need for excessive thinking or learning; you can just take it out and start shooting.

However, the Pocket 3 is not perfect yet. It has a single focal length, and data transfer still requires using an app, making it less seamless. These imperfections are precisely where the advantages and opportunities lie for mobile phone manufacturers.

Today, the DJI Pocket has been around for eight years. With the Pocket 4 still shrouded in mystery and smartphone manufacturers launching a collective attack, this young and interesting form factor has finally ushered in its "Normandy moment".

#Welcome to follow iFanr's official WeChat account: iFanr (WeChat ID: ifanr), where more exciting content will be presented to you as soon as possible.

ifanr | Original Link · View Comments · Sina Weibo

The Xiaomi YU7 GT was officially unveiled in the 404th batch of the "Announcement of Road Motor Vehicle Manufacturers and Products" issued by the Ministry of Industry and Information Technology.

The new car measures 5015 mm in length, 2007 mm in width, and 1597 mm in height, with the wheelbase remaining unchanged at 3000 mm.

Compared to the current YU7 model (4999mm long, 1996mm wide, and 1600mm high), the new car is slightly longer and wider, while the height is slightly lower, resulting in a lower overall stance and a significantly enhanced sense of sportiness.

In terms of body details, the YU7 GT has eliminated the front and rear active air intake grille (AGS) system, while making minor adjustments to the side skirts and wheel arches, and adding several aerodynamic components, making the overall visual effect sharper and more track-oriented.

In addition, a striking red "GT" logo has been added to the front door panel connection and the rear of the car to further emphasize its high-performance identity.

It is worth noting that the model submitted this time uses black paint, a color that has not appeared on the regular YU7 before. Black may become an exclusive or debut color for the GT version.

The powertrain is the core highlight of the YU7 GT. The new car is equipped with a high-performance dual-motor system jointly developed by Xiaomi Auto and Inovance Technology. The front axle motor is the same as that in the SU7 Ultra, with a power of 288 kW; the rear axle is equipped with a single V8 motor with a peak power of 450 kW; the combined output power is approximately 1004 horsepower, and the top speed can reach 300 km/h.

In comparison, the current YU7 Max has front and rear motors with power outputs of 220 kW and 288 kW respectively, for a total power output of 508 kW (approximately 690 hp). The YU7 GT represents a near doubling of its performance.

To match its high-performance positioning, the vehicle's tire specifications have been upgraded accordingly, with the front tires being 265/40 R21 and the rear tires being 295/35 R21, both being 21-inch ultra-wide tires, providing stronger grip and handling stability.

The YU7 GT also offers a wide range of optional extras, including aerodynamic or visual kits such as front combination light trim, carbon fiber exterior mirrors, rear spoiler, ducktail spoiler, and hood trim.

According to multiple sources, the YU7 GT has entered the real-vehicle testing phase and has been spotted multiple times at the Nürburgring Nordschleife for intensive tuning.

The new vehicle will be equipped with an active hydraulic suspension system that can automatically lower the vehicle height in track mode and support free adjustment of power distribution, aiming directly at the lap record for mass-produced SUVs, fully demonstrating its engineering philosophy of "born for the track".

The YU7 GT is primarily positioned to compete with high-performance SUVs such as the Tesla Model Y Performance and the Porsche Cayenne Turbo GT, targeting consumers who frequently participate in track events and seek the ultimate driving experience.

Despite its powerful performance, judging from its overall specifications, its overall positioning is still slightly lower than Xiaomi's flagship model, the SU7 Ultra. The new car is expected to be priced between 400,000 and 500,000 yuan.

Xiaomi previously established its "performance technology" brand image with the SU7 Ultra, and the launch of the YU7 GT will further solidify this strategic label and expand its layout in the high-performance electric product matrix.

In addition to the YU7 GT, Xiaomi plans to launch the next-generation SU7 in April this year, and has already planned to include more new models, including the YU9, a large and mid-sized extended-range family SUV, and the SU7 L, a four-door executive sedan.

Since the delivery of its first model on April 3, 2024, Xiaomi Auto has delivered more than 500,000 vehicles and achieved profitability in the third quarter of 2025.

In a public live broadcast in January this year, Xiaomi founder Lei Jun further raised the annual delivery target for 2026 to 550,000 vehicles.

#Welcome to follow iFanr's official WeChat account: iFanr (WeChat ID: ifanr), where more exciting content will be presented to you as soon as possible.

ifanr | Original Link · View Comments · Sina Weibo

Just now, a "collision of Mars and Earth" occurred in the Silicon Valley AI circle.

As if by prior arrangement, OpenAI and Anthropic simultaneously released their major updates: Claude Opus 4.6 and GPT-5.3-Codex.

If we were discussing "how to write good prompts to assist with work" until last night, then this morning we may have been forced to learn "how to manage AI employees as a boss".

AI creates AI, and incidentally takes over your computer.

Just yesterday, Sam Altman celebrated Codex's "one million active users" milestone on the X platform. A mere day later, OpenAI followed up with another bombshell – GPT-5.3-Codex.

The technical documentation contains a very significant statement: "This is the first model that played a key role in our creation process."

In layman's terms, this means that AI has learned to write its own code, find bugs on its own, and even begun training the next generation of AI. This self-evolutionary capability is directly reflected in a series of benchmark scores.

Remember the OSWorld-Verified benchmark test that simulates human operation of a computer? The previous model only had an accuracy of 38.2%, which was not even up to the passing mark.

But this time, GPT-5.3-Codex jumped to 64.7%!

It's worth noting that the average human level is only 72%. This means that AI is just a hair's breadth away from being as adept as you are at using a mouse, switching screens, and operating software.

In Terminal-Bench 2.0 (command line operation), it achieved a high score of 77.3%, far surpassing GPT-5.2 (62.2%).

The renowned SWE-Bench Pro benchmark covers four programming languages, is not only resistant to contamination, but also tackles real-world, hardcore engineering challenges.

GPT-5.3-Codex not only achieved state-of-the-art (SOTA) performance here, but also used fewer tokens than any previous model. What does this mean? It means that it not only works incredibly fast, but also solves problems in a shorter and more cost-effective way than humans.

OpenAI even demonstrated its ability to build independently:

Within days, it built a racing game v2 with multiple maps from scratch, and also managed to create a deep-sea diving game that manages an oxygen system.

What impressed me most was GPT-5.3-Codex's understanding of ambiguous intent.

When building the "Quiet KPI" landing page, it automatically converted the "annual plan" into the "discounted monthly price" and even thoughtfully added a user review carousel—all without you having to give any instructions.

OpenAI's ambitions are written all over its face: Microsoft used to say that AI would become the copilot of humans, but now AI wants to be the driver who can take control of the steering wheel and even fix the car itself.

Oh, and there's one more interesting detail.

Previously, it was widely rumored that OpenAI had reservations about NVIDIA's AI chips, but this time the official blog specifically emphasized that the design, training and deployment of GPT-5.3-Codex were all completed on the NVIDIA GB200 NVL72 system.

This high-EQ "thank you Nvidia" really gave Huang Renxun a lot of face.

Saying goodbye to "goldfish memories," Claude staged a dramatic comeback.

Around the same time as the release of GPT-5.3-Codex, Anthropic also presented its own Chinese New Year gift package.

The bad news is that the much-anticipated "medium" Claude Sonnet model hasn't been updated; but the good news is that Anthropic has directly presented the "super-sized" version – Claude Opus 4.6.

Compared to OpenAI's aggressive approach to "action," Anthropic's Claude Opus 4.6, released today, focuses on "thinking ability" and "usability."

Many enterprise users have a pain point called Context Rot: it claims to support 200k contexts, but when a lot of data is put in, the AI ​​starts to focus on the beginning but not the end.

This time, the data presented by Claude Opus 4.6 is simply a "game-changer".

In the MRCR v2 (Long Text Needle in a Haystack) test, Claude Opus 4.6 achieved a recall rate of 76%.

In contrast, the previous generation Sonnet 4.5 had a dismal 18.5%. In a sense, this represents a qualitative leap from being virtually unusable to being "highly reliable".

This is the first time that Claude Opus 4.6 has introduced a truly usable 1M context window.

What does this mean? It means you can throw hundreds of pages of financial reports or hundreds of thousands of words of code directly at it, and it can not only read them all, but also accurately tell you that there is a problem with the number in the footnote on page 342.

What really caught the eye of working people was its productivity function.

On one hand, Anthropic has now integrated Claude directly into Excel and PowerPoint. It can generate PPTs directly from Excel data, preserving not only the layout style but also aligning fonts and templates. In the Claude Cowork collaboration environment, it can even perform autonomous multitasking.

On the other hand, Anthropic took the opportunity to launch an experimental Agent Teams feature in Claude Code, allowing ordinary developers to experience the feeling of "commanding thousands of troops":

• Role division: You can designate a Claude Session as the Team Lead, which does not do the dirty work and is specifically responsible for breaking down tasks, assigning work orders, and merging code; the other Sessions are teammates (Teammates), each taking on tasks to do.
• Independent operation: Each teammate has an independent context window (no need to worry about token explosion), and they can even send messages to each other behind your back (inter-agent messaging) to discuss technical details, and finally only report the results to the team leader.
• Parallel horse racing: What's the use of this? Imagine checking for a stubborn bug. You can generate 5 agents to verify 5 different hypotheses, like a "horse race" to clear the mine in parallel; or during code review, you can have one teammate act as a "security expert" to check for vulnerabilities, and another act as an "architect" to check performance, without interfering with each other.

To demonstrate the limits of Opus 4.6, Anthropic researcher Nicholas Carlini conducted a crazy experiment: Agent Teams.

Instead of writing the code himself, he threw out $20,000 in API credits, allowing 16 Claude Opus 4.6 users to form a "fully automated software development team".

In just two weeks, this group of AIs autonomously conducted more than 2,000 programming sessions and wrote a C language compiler (based on Rust) with 100,000 lines of code from scratch.

This AI-written compiler also successfully compiled the Linux 6.9 kernel (covering x86, ARM, and RISC-V architectures) and even ran the Doom game.

While it's not perfect (for example, the generated code isn't as efficient as GCC), this case demonstrates that we're no longer programming with AI, but rather watching an AI team autonomously collaborate, debug, and advance the project.

In addition, it has learned Adaptive Thinking, which allows it to decide "how long to think" based on the difficulty level. With the addition of a new "intelligent intensity" control, you can switch between four levels, from Low to Max.

In terms of pricing, Anthropic has been quite generous this time, maintaining a base price of $5/$25 per million tokens. It seems determined to compete head-to-head with OpenAI in the enterprise market.

One is a radical genius, the other is a reliable old cow.

Renowned AI reviewer Dan Shipper conducted a blind test (Vibe Check) immediately, and his evaluation was remarkably accurate:

Claude Opus 4.6 is characterized by "High Ceiling, High Variance".

It's like a brilliant but occasionally eccentric genius. In testing, it directly solved a feature problem that had stumped the iOS team for two months; it achieved a high score of 9.25/10 in the LFG Benchmark.

But it can also be "overconfident" at times, spouting nonsense with a straight face. If you need a breakthrough inspiration, choose it.

GPT-5.3-Codex is "High Reliability, Low Variance".

It's like a seasoned, reliable engineer who never lets you down. Reasoning speed is improved by 25%, it makes almost no basic mistakes, and its stability is reassuring.

While slightly less efficient at creative tasks (LFG score 7.5/10), it's the most efficient workhorse for daily coding and operations. Choose it if you need stable delivery.

As we move into 2026, our roles begin to change.

At this juncture, the biggest change for ordinary users is this: the importance of Prompt Engineering is declining, while the capabilities of Agent Management are beginning to emerge.

When ChatGPT can fix bugs and even operate your terminal autonomously, and when Claude can process 1 million words at a time and pinpoint details precisely, we no longer need to break down commands into fragmented instructions like teaching a primary school student.

What we need to do is learn how to define goals, review results, and—decide when and which task to assign to which "employee" in the role of a "manager".

This is the new workplace in 2026: your team is infiltrated by a group of silicon-based geniuses, and you are the only carbon-based boss.

#Welcome to follow iFanr's official WeChat account: iFanr (WeChat ID: ifanr), where more exciting content will be presented to you as soon as possible.

ifanr | Original Link · View Comments · Sina Weibo