ChatGPT may add a $500-per-month Pro Max subscription plan
要闻
Unreleased ChatGPT subscription information suggests that OpenAI may add a Pro Max plan priced at $500 per month. Some pages show a $600 price, which reports attribute to value-added tax, while the final benefits and launch timing remain undetermined.
OpenAI’s unreleased ChatGPT subscription information includes a Pro Max plan priced at $500 per month. The plan is not currently available, and users cannot subscribe to it. Another page shows a price of $600, with reports explaining that the difference between the two prices comes from value-added tax. Compared with the existing Pro plan, the main wording change found is the addition of “fastest Work and Codex,” indicating faster processing for related tasks, although specific speeds and usage allowances have not been stated. Media reports have also speculated that the plan may use Cerebras computing infrastructure, but OpenAI has not confirmed a direct connection between the two, and neither the final benefits nor the launch timing has been announced.
Doubao offers another 30-day free subscription for desktop users
要闻
Following its previous 30-day subscription offer, Doubao is providing another 30 days free to users of the latest Doubao and Doubao Work desktop apps. Each account can claim the offer once, while recipients of the previous offer will receive the next one as their current benefit nears expiration.
Effective immediately, Doubao is offering another 30-day free subscription on top of its previous one-month subscription campaign. The offer is available to users who download or upgrade to the latest version of the Doubao or Doubao Work desktop app, and each account can claim it only once. According to the company, users who received the previous 30-day benefit will receive the next 30-day subscription as the existing one approaches expiration. For users who already had a paid subscription before claiming the offer, the paid subscription period will be extended automatically, while benefits such as priority access during peak hours will remain available throughout. Users can check the expiration date under “Settings—Subscription and Quota Management” and view further details through “Subscription History” on the same page.
Google adds Colab premium benefits to Google AI subscription plans
要闻
Google has added premium Colab benefits to Google AI subscriptions, giving users priority access to faster accelerators and more powerful machines. Ultra subscribers also receive uninterrupted background execution and Premium GPU access, with the benefits rolling out to Colab-supported countries over the next few weeks.
Google announced that Google AI subscribers can now receive premium Colab benefits, with the related features rolling out over the next few weeks to subscribers in countries where Colab is supported. Google AI subscriptions provide priority access to faster accelerators and more powerful machines; Google AI Ultra also includes uninterrupted background execution and Premium GPU access, allowing long-running training jobs to continue without keeping a browser tab open. According to the company, Google AI plans also include expanded cloud storage, broader access to Google’s most capable models, enhanced features in the Gemini app, Workspace integrations, and premium access to developer tools including Google Antigravity and Google AI Studio. Existing Colab subscriptions will not change, and their benefits can be stacked with Google AI subscription benefits to provide additional access to premium Colab features.
WorkBuddy launches WeChat Mini Program generation and publishing capabilities
要闻
WorkBuddy has added WeChat Mini Program generation and publishing to PC version 5.6.1 and later. Users can create projects with natural-language instructions, connect cloud services, preview projects, and start publishing within the product, while those without an account can create a 14-day trial version.
WorkBuddy has launched WeChat Mini Program generation and publishing in PC version 5.6.1 and later, with the features available after upgrading to an eligible version. According to its description, users can select “Mini Program” in “Code Development” mode or describe their requirements directly in natural language. After a project is generated, they can connect cloud services such as cloud databases, file storage, login authentication, and AI calls as needed, then preview it and enter the publishing process within WorkBuddy. The standard process does not require a separate developer tool, AppID configuration, or manual uploading of a code package. Users with an existing Mini Program account can bind it by scanning a QR code, publish a test version for testing on a physical device, and then submit it for review and release the production version through WeChat’s official process. Users without an account can create a 14-day trial version to validate functionality, but it cannot be converted directly into a production version.
A DeepSeek Harness Desktop installer has appeared on DeepSeek AI’s GitHub Releases page. The page also lists multiple changelog ranges for the 0.1.6 and 0.1.7 series.
A DeepSeek Harness Desktop installer appeared on DeepSeek AI’s GitHub Releases page, but the source does not specify when it was discovered or released. The page shows Full Changelog ranges from dsh-v0.1.7-rc.1 to dsh-v0.1.7-rc.2, from dsh-v0.1.7-alpha.1 to dsh-v0.1.7-alpha.2, and from dsh-v0.1.6-alpha.1 to dsh-v0.1.6-alpha.2. According to the page, the first release candidate in the 0.1.7 series consolidates the main user- and developer-related changes since v0.1.5-rc.3. The original page displays multiple loading errors, and the provided content does not include the installer filename, supported platforms, or installation instructions.
SiliconFlow launches APIs for three open-source fast decision-making models, free through October 8
开发生态
SiliconFlow has launched Serverless APIs for three open-source models—Kev-4B, SemIf, and DiffusionGemma. Developers can use them free of charge before October 8, 2026, for fast structured judgments such as yes-or-no, single-choice, and rating-scale assessments.
SiliconFlow has deployed Kev-4B, SemIf, and DiffusionGemma as Serverless services accessible through APIs, with free calls available to developers before October 8, 2026. To use them, developers must register or sign in to the SiliconFlow platform and obtain an API Key. Each request can specify a model, provide content to evaluate, and set the question format to yes-or-no, single-choice, or rating-scale assessment, while multiple question types can be combined in one request. According to SiliconFlow, the three open-source fast-decision models use different technical approaches and target applications that require structured judgments quickly.
Bocha releases the Bocha Jev decision-making model with limited-time free API access
开发生态
Bocha has released Bocha Jev, a model for structured decision-making, and opened a limited-time free API trial. bocha-jev-v1 handles up to 32 questions and 1,024 candidates per request, with a default limit of 200 QPS per user.
Bocha has released Bocha Jev for structured decision-making tasks and opened a limited-time free API trial, which developers can access with an existing Bocha API Key. According to the company, an application can submit the current state, task instructions, and candidates, then use the model’s action selection, content scoring, or condition evaluation results to perform subsequent operations. The model supports multiple questions sharing the same context in one request and can be used for customer-service routing, search-result ranking, and Agent workflows. The current model identifier is bocha-jev-v1; each request supports up to 32 questions and 1,024 candidates, while the default request limit is 200 QPS per user, subject to the account configuration.
Anthropic resumes charging for three categories of requests blocked before a response
开发生态
Anthropic bills requests refused before output in the bio, frontier_llm, or reasoning_extraction category at the model’s rate. Other categories and null are not billed, but every request still counts against rate limits.
Anthropic’s documentation states that when the safety classifier in Claude Fable 5.1, Claude Fable 5, Claude Opus 5.5, or Claude Opus 5 refuses a request before output under the bio, frontier_llm, or reasoning_extraction category, the request is billed at the normal rate of the model that ran it; other categories and null are not billed. According to the company, these three categories had low volumes of false positives as of September 2026. The rules cover the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry. A refusal still returns HTTP 200, with "refusal" in stop_reason and the category in stop_details.category. Whether billed or not, usage shows the token counts, and the request counts against rate limits. If a refusal occurs during streaming, the input tokens and tokens already generated are billed at normal rates, and the existing output must be treated as incomplete and discarded.
Server-side fallback for the Claude API is in beta. When a request includes the server-side-fallback-2026-07-01 beta header and sets fallbacks to "default," the API passes the same request within the same call to Anthropic’s recommended model when the refusal category has a recommended fallback; if no model is recommended, the refusal remains. Developers can instead specify up to 3 fallback models and retry them in order. Only a safety classifier refusal triggers this mechanism, while a rate limit, overload, or server error is returned unchanged. If the original refusal that triggers fallback belongs to one of the billed categories or occurs mid-stream, it is billed separately from the fallback request. Fallback credit compensates for the later request’s prompt-cache miss so that cache creation for the same conversation is not charged twice. The top-level model field identifies the model that answered, and fallback_message in usage.iterations records the handoff.
Claude Code Projects adds local support, allowing threads to run on users’ computers
开发生态
ClaudeDevs announced that Claude Code Projects now supports local operation, allowing project threads to run on a user’s own computer. The feature is being tested by a limited group on desktop and web and is gradually expanding to more Pro and Max users through a waitlist.
ClaudeDevs announced that Claude Code Projects now supports local execution, allowing project threads to run on a user’s own computer, with the feature currently in limited testing on desktop and web. According to the official statement, the team is expanding access through the waitlist to grant availability to more Pro and Max users. People who previously used Claude Projects are still waiting to be migrated, while those who have not yet registered can apply for early access.
OpenRouter launches a server-side tool marketplace offering search, command-line tools, and more
开发生态
OpenRouter has launched a server tool marketplace listing 11 capabilities, including Web search, URL content extraction, sandboxed shells, image generation, multi-model calls, and on-demand loading of tool definitions.
OpenRouter has launched a server tool marketplace whose current page lists 11 server tools for use by models. Information-access capabilities include providing the current date and time in UTC or a configured timezone, searching the Web, and fetching and extracting content from URLs. Execution capabilities include a hosted sandboxed shell for any model and an Anthropic-style bash tool with optional sandboxed execution. Model orchestration covers consulting a stronger model during generation, calling multiple models and using another model to synthesize the final answer, and delegating tasks to a smaller, faster model. The remaining capabilities include generating images from text prompts, having models propose file edits as V4A diffs, and loading tool definitions only when needed.
Google officially launches Gemini 3.8 Live’s video avatar feature on its enterprise platform
开发生态
Google has released Gemini 3.8 Live with Live Avatar in Gemini Enterprise, combining real-time voice conversations with low-latency video avatars and supporting switching among 97 languages.
Google made Gemini 3.8 Live with Live Avatar available in Gemini Enterprise on the announcement date, combining near-real-time video generation with native live voice dialogue. According to the company, the feature can process visual and audio inputs simultaneously and respond with audio and video that include lip-syncing, facial expressions, and fluid turn-taking. It can be used for customer service and interactive walkthroughs. Its asynchronous tool calling can trigger tools and retrieve data in the background while the conversation continues, supporting tasks such as checking a guest into a hotel.
According to Google, Live Avatar includes native multilingual speech-to-speech synchronization, can switch among 97 languages, and dynamically adjusts lip movements and expressions. The company says this process does not reduce video fidelity or introduce visual drift. Organizations can use preset Avatars or generate custom Avatars from a high-quality reference image while preserving a person’s likeness, brand styling, or character identity, but this capability is currently limited to customers on the enterprise allowlist. Google says all output generated by its AI products carries a SynthID watermark, including an imperceptible marker embedded in Live Avatar audio and video. API documentation and a model card are also available through the product page.
Claude Code plans to turn plan mode into a built-in mod
开发生态
Claude Code developer Thariq plans to turn plan mode into a built-in mod and let other mods add new modes or reassign Shift+Tab. The proposal remains in planning and has not shipped.
Claude Code developer Thariq said he plans to turn plan mode into a built-in mod and allow other mods to add new modes or reassign Shift+Tab; the change has not shipped. Under the proposal, users could edit the plan mode prompt, create and share custom modes, or forgo plan mode and bind the shortcut to another function. Thariq had previously considered removing plan mode and using Shift+Tab to adjust how much thinking effort the model applies. After seeking feedback, he said some users plan tasks themselves and do not need a dedicated mode, while others want a mode used only to discuss and develop ideas with Claude.
Odyssey releases Agora-2, enabling up to 20 humans and Agents to share a real-time simulated world
模型发布
Odyssey released a playable research preview of Agora-2 on September 21, 2026, allowing up to 20 humans and Agents to enter one real-time generated environment. Its participant capacity is 5 times that of Agora-1, and it expands from one environment to multiple environments.
On September 21, 2026, Odyssey released a playable research preview of Agora-2, its next-generation multi-agent world model, supporting real-time interaction among up to 20 humans and Agents in a shared simulation. According to the company, its participant capacity is 5 times that of Agora-1, while its simulation scope has expanded from one environment to multiple environments to cover more complex interactions and longer-horizon behaviors. Agora-2 was trained on captures from Diablo II, developed by Blizzard, pairing observation with action and state so the model can learn how entities move, interact, and respond to one another.
According to Odyssey, Agora-2 generates a shared world as streaming pixels and maintains an explicit shared state. The simulation model uses entity properties, recent action, and surrounding geometry to predict how combined action affects that state. After the world server combines the predictions and updates the state, the rendering model generates each participant’s view. When an entity moves out of frame, its properties remain in the shared state. The rendering model is trained with flow matching; during training, visual history is frequently removed, and greater weight is assigned to errors involving entities. Odyssey also uses reinforcement learning to train Agents to pursue opponents, navigate around obstacles, and recover after becoming stuck or separated. These Agents act using a partially observed view and recent observation.
Tencent launches the Hy translation app with support for translation across 33 languages
产品应用
Tencent has launched the Tencent Hy Translation app, built on its Hunyuan Hy-MT2 translation model, in 12 countries and regions including the United States and Japan. The app supports translation across 33 languages, as well as voice, photo, and offline translation.
Tencent has released the Tencent Hy Translation app, built on its Hunyuan Hy-MT2 translation model, in 12 countries and regions including the United States and Japan. According to the company, the app supports translation across 33 languages, covering 5 Chinese ethnic minority languages and dialects, and offers voice, photo, and offline translation. Users can download the translation model to their phones in advance and use the offline function without an internet connection or when reception is poor. In addition to the mobile app, Tencent provides a mini program, plugins, and a PC version, which users can access by searching for “Tencent Hunyuan Translation.”
Arena.ai launches a redesigned leaderboard overview
产品应用
Arena.ai has consolidated live leaderboards, a new-model feed, and platform updates into its redesigned leaderboard overview. On one page, users can view scores, category rankings, and capability summaries for models including GPT-6 Sol and Claude Opus 5.5.
Arena.ai has launched a redesigned leaderboard overview that brings previously scattered real-time model information onto one page for researchers, developers, and model-evaluation participants to track performance. The new page includes a release feed for models recently added to Arena and displays live evaluation counts along with leading models in categories such as Agents, image generation, and coding. It also features the Arena team’s preliminary descriptions of new-model capabilities and updates on the platform’s research, features, and releases. Users can view the scores of models including GPT-6 Sol and Claude Opus 5.5 across different leaderboards, then assess their performance alongside category rankings and capability information.
Tencent QClaw to shut down on December 24, with refunds and data migration available
产品应用
Tencent’s QClaw will cease operations at 00:00 on December 24, 2026, and has already stopped new-user registration, subscription purchases, renewals, and automatic billing. Users can request refunds or migrate data to WorkBuddy, with 1,000 points offered for successfully migrating valid data.
The QClaw team announced on September 24, 2026, that QClaw will cease operations at 00:00 on December 24, 2026, citing adjustments to business development. From the announcement date, new-user registration, subscription purchases, renewals, and automatic renewal charges were stopped; registered users can continue using the service until shutdown, and existing points retain their original expiration dates. Users can request refunds through the official website for unused, still-valid paid subscriptions or point packages, or back up their data in the client. Migration to WorkBuddy requires authorization on the official website followed by synchronization through the client. After shutdown, only the data backup and migration functions will remain available. Users who successfully migrate valid data will receive 1,000 WorkBuddy points, while eligible users purchasing a one-month membership for the first time will receive additional points.
Meta moves to integrate Muse with Mac, email, and more shopping services
产品应用
Meta is expanding its personal AI Agent, Muse, to video conversations, smart glasses, Mac, email, and more shopping services, while opening a platform for developers to build their own connectors. Smart-glasses integration is planned for release in the coming months.
Meta announced an expansion plan for its personal AI Agent, Muse, at Connect, extending its scope to digital avatars, smart glasses, Mac operations, email collaboration, and shopping services. Smart-glasses integration is planned for release in the coming months. According to the company, users will eventually be able to hold real-time video conversations with Muse’s digital avatar and assign tasks by voice through Meta smart glasses. Meta also presented Muse’s ability to operate Mac applications on a user’s behalf and said it is developing a dedicated mailbox for the Agent to handle email. For shopping, Muse will connect to more retail applications and payment methods. The company said Muse will provide a large number of free tokens and is expected to charge small transaction fees over the long term. Meta has also opened a platform that lets developers build their own connectors to expand the services available to Muse.
Meta unveils the Muse Charm wearable, with shipments planned before the December holidays
产品应用
Meta’s Muse AI Agent is getting a keychain-style device called Muse Charm. The company unveiled it on September 23, 2026, and aims to ship it before the December holidays, although the internal component layout still needs to be finalized.
Meta CEO Mark Zuckerberg presented Muse Charm at the company’s annual Connect event on September 23, 2026, and said the device would be ready to ship in time for the December holidays. According to Zuckerberg, the team still needed to finalize the internal component layout, so the product was not ready for delivery at the time. Muse Charm is palm-sized and can be carried in a pocket or attached to a keychain. Its small screen displays a digital avatar known internally as Jolly, and users can interact with Muse by voice. According to the report, the device is designed to carry out tasks on the user’s behalf. Zuckerberg said Meta had put Muse’s real-time voice and avatar stack into the device. Touching the fingerprint sensor in the corner starts a conversation without requiring the user to unlock a phone or open an app. He also said that, when users are not wearing glasses, it would be the fastest way to talk to Muse and show it what is happening around them.
Meta unveils three smart glasses, with VR glasses planned for spring 2027
产品应用
Meta unveiled three smart glasses: the $1,299 Meta VR Glasses are scheduled for spring 2027, while the other two Ray-Ban Meta models are priced at $349 and $449.
At Meta Connect, Meta introduced the $349 Ray-Ban Meta Audio and the $449 third-generation Ray-Ban Meta, while announcing that the $1,299 Meta VR Glasses will go on sale in spring 2027. According to the company, the VR model weighs one-fifth as much as the Meta Quest 3 and has a 5K display. Its computing components, battery, and storage sit in a separate wired puck, while external cameras show the surrounding environment. Users can control the device through hand gestures or the Muse AI agent; it can create one or multiple virtual displays, map flat surfaces into virtual keyboards and mice, and show a fully rendered custom Muse avatar. Meta also says it is the first IMAX Enhanced-certified VR device. It will support Disney+, with Prime Video, YouTube, and other services planned, while Xbox Cloud Gaming is also coming to the device. The Meta AI app can create a full-body digital representation from a user’s photo for video calls while wearing the glasses.
Ray-Ban Meta Audio removes the cameras, provides up to 12 hours of battery life, and supports music playback and conversations with the Muse AI agent. Existing camera-equipped models use an LED to indicate recording, but some owners have tampered with the light, and the resulting concerns have led some cities, states, and businesses worldwide to restrict camera-equipped smart glasses in certain settings. The third-generation Ray-Ban Meta retains its cameras while updating battery life and microphones. Meta Glasses, a separate line from Ray-Ban, were also updated and start at $249. IDC data shows that Meta accounted for 68.7% of global smart-glasses shipments in the second quarter of 2026. Samsung and Google are preparing glasses based on Android XR with designs from Warby Parker and Gentle Monster. Snap is rolling out the $2,195 Specs AR, while Bloomberg’s Mark Gurman reported that Apple may debut its own smart glasses in late 2027.
Google unveils a multi-Agent research framework for long-form video generation
技术与洞察
Google has unveiled a multi-Agent approach for long-form video that spans 4 frameworks and uses 2 loops to coordinate creative choices and production. The company says it can generate minutes of content while reducing cross-shot character changes and the propagation of upstream errors.
Google has published an AI video co-director multi-Agent research framework for coherent long-form video generation. Built as an orchestration layer on Gemini and Veo, it models multi-shot storytelling as a global optimization and world-state tracking problem. The research includes Co-Director, CANVAS, A²RD, and VQQA; Co-Director is slated to appear at COLM 2026, while CANVAS is slated to appear at EMNLP 2026. According to Google, generated images, video, and audio inherit safety mechanisms from the foundation models, including SynthID watermarking, and a safety classifier can also be applied to the finished video.
The system operates through 2 loops: strategic steering and multi-stage production. The Orchestrator Agent uses a multi-armed bandit to select configurations across 3 dimensions: Creative Strategy, Narrative Mode, and Aesthetic Archetype. The Pre-Production Agent combines a scene-by-scene storyline, visual assets, and a storyboard, after which the Production Agent uses the Keyframe Agent, Video Agent, and Audio Agent to create visuals, motion, voiceover, and music. Finally, an MLLM Judge evaluates the completed video across the same 3 dimensions and returns a factored reward signal to the MAB for iterative optimization. Google says its evaluations showed improvements in multi-shot narrative consistency and character persistence, with the system generating minutes-long videos while mitigating visual drift and pipeline error propagation.
Anthropic introduces Project Swap, in which Claude Agent exchanges books on behalf of employees
技术与洞察
Anthropic said Project Swap had 201 employees each contribute one book for exchange while Claude Agents brokered the trades. The experiment ran across six offices and evaluated the results using participants’ ranked book preferences.
According to Anthropic, the company used Project Swap during the summer to organize a book exchange among 201 employees. Each participant contributed one book they wanted to give away, discussed their reading preferences with a Claude-powered Agent, and then had the Agent search for trading partners on a digital trading floor until the allotted time expired. Participants were divided into six local pools in San Francisco, New York City, London, Seattle, DC, and Dublin, ranging from 3 people in Dublin to 115 in San Francisco. Most participants received a book acquired by their Agent, while the Dublin participants were excluded from most of the analysis.
To evaluate whether trades matched individual preferences, Claude first used a short, semi-structured conversation to ask about reading tastes and what each participant wanted to read during the summer. Fable 5 then ranked every book in that participant’s pool. Separately, the research team asked each person to rank 10 books from their pool as ground truth. Anthropic said the experiment examined whether Agents could accurately understand participants and how Agent-oriented markets should define admission, failed-trade handling, and the visibility of market activity. Unlike the earlier Project Deal, in which Agents bought and sold varied physical goods, this experiment used book-preference rankings to calculate how well the market satisfied participants’ preferences.
Google partners with Planet to launch a satellite and test TPU operation in space
行业动态
Google will work with Planet to send a TPU-equipped prototype satellite into orbit on SpaceX’s Transporter-18 mission, testing whether the chips can operate under acceleration of up to 100 g, radiation and vacuum-cooling conditions. A two-satellite laser communication test is scheduled for 2027.
Google has decided to work through Project Suncatcher with Planet to launch a prototype satellite on SpaceX’s Transporter-18 rideshare mission, collect in-orbit data on TPUs exposed to flight vibration, radiation and thermal extremes, and set 2027 as the next-stage milestone. The project is examining whether low Earth orbit can host scalable machine learning infrastructure. According to the company, satellites can use near-constant sunlight to generate up to eight times as much solar power as on Earth. A rocket trip to low Earth orbit takes about 10 minutes, with spacecraft experiencing sustained acceleration of up to 10 g and individual components such as TPUs potentially facing 50 to 100 g. The team conducted vibration tests along all three axes, and the company said the hardware withstood the forces applied during testing.
The team also used a proton beam at UC Davis’s Crocker Nuclear Laboratory while running AI workloads on Trillium TPUs and monitoring errors such as bitflips. According to the company, initial results showed that the chips could withstand a total ionizing dose greater than the dose expected during a five-year space mission. To address the lack of airflow in a vacuum, the project uses a combination of heat pipes and radiators to cool the chips and has tested the system in a thermal vacuum chamber. Future satellite designs will each carry dozens of TPUs and orbit Earth in clusters. The satellites will use lasers to determine their own positions and their positions relative to neighboring satellites, maintaining very-high-bandwidth connections over extremely short distances. The project plans to place two satellites in orbit in 2027 to test this approach.
DeepSeek reportedly pursuing a new funding round, with over 70% of its compute used for model training
前瞻与传闻
DeepSeek is reportedly pursuing a new funding round of about RMB 50 billion at a target valuation of roughly RMB 500 billion, with completion planned by the end of October 2026 at the latest. Its annualized revenue run rate has reached $1 billion, while more than 70% of its computing capacity is allocated to training new models.
DeepSeek is reportedly finalizing a funding round of about RMB 50 billion at a target valuation of roughly RMB 500 billion, with plans to complete it by the end of October 2026 at the latest. According to the report, the company’s annualized revenue run rate is $1 billion, more than double the level of less than $500 million recorded several months earlier. Founder Liang Wenfeng reportedly told an investor meeting that customer numbers had not declined after model prices were raised, but increasing revenue was not the current priority. The report said the company allocates more than 70% of its computing capacity to training new models and less than 30% to inference services for existing models.
Musk expects SpaceX could have a Fable- or GPT-6-level model within two to three months
前瞻与传闻
Musk said he is cautiously optimistic that SpaceX could have a model at the level of Fable or GPT-6 within two to three months. He also predicted that the company could take the lead in about six months if development keeps accelerating, and said SpaceX has demonstrated an ability to deploy large-scale compute rapidly.
Musk said in a social media post that SpaceX could have a model at the level of Fable or GPT-6 within two to three months and, if progress in AI R&D continues to accelerate, could be in the lead in about six months. According to his account, SpaceX will continue accelerating its AI R&D; its AI operation has been underway for about three years, compared with about six years for Anthropic and about ten years for OpenAI. Musk also said that rapidly deploying large-scale compute is extremely difficult and that SpaceX has already demonstrated this capability.
Google reveals Gemini 4 progress: in early post-training and already used in internal coding tools
前瞻与传闻
Google DeepMind’s head said its next flagship model, Gemini 4, has entered early post-training. Related reporting also said engineers are using it internally to run the Antigravity AI coding tool. Google plans an early release well before year-end, but no launch date has been set.
Google DeepMind head Koray Kavukcuoglu said at The Information’s AI Agenda Live summit that Gemini 4 has entered early post-training and that Google plans to release an early version well before year-end, followed by rapid iteration. The timing represents a release intention rather than a confirmed launch date. According to related reporting, Google is also developing safeguards and conducting safety tests for Gemini 4, while its engineers are already using the model internally to run the Antigravity AI coding tool. External availability remains dependent on the planned early release.
Anthropic reportedly plans to adjust its equity structure, giving seven co-founders 50.1% of voting power
前瞻与传闻
Citing people familiar with the matter, The Information reported that Anthropic plans to restructure voting rights before its IPO, giving seven co-founders, including Dario Amodei, a combined 50.1% vote on most corporate matters.
Citing people familiar with the matter, The Information reported that Anthropic is seeking shareholder approval for a pre-IPO equity restructuring that would grant a special class of shares to CEO Dario Amodei and six other co-founders, giving the seven a combined 50.1% of the voting power on most corporate matters. The arrangement would require at least three of the seven to continue holding a specified amount of company stock, but it would not apply to board elections. The report also said Anthropic plans to grant employees another class of special shares that would determine the outcome when votes on certain corporate matters are tied.
Google, OpenAI, and Anthropic reportedly plan to form a self-regulatory organization for AI safety standards
前瞻与传闻
Google, OpenAI, and Anthropic are reportedly preparing a self-regulatory body tentatively called SAFA to set frontier AI safety standards without government oversight, with a possible launch in late 2026 or early 2027.
According to The Information, Google, OpenAI, and Anthropic are working to establish a self-regulatory organization for AI safety standards aimed at frontier model developers, with no government oversight planned and a possible launch in late 2026 or early 2027. People familiar with the matter said the group is tentatively named the Standards Authority for Frontier AI (SAFA). It would support third-party testing before models are deployed, define procedures for reporting safety and security incidents, and set qualification requirements for independent auditors. The working group is still discussing whether the organization will directly conduct model safety and capability testing, so its responsibilities have not been fully determined. The project remains in the planning and discussion stage.
