OpenAI has released GPT-6 Astra to a small number of institutions, with paid ChatGPT users, the OpenAI API, and AWS set to follow over the next few days. Standard API pricing is $10 per million input tokens and $50 per million output tokens.
OpenAI has begun a phased release of GPT-6 Astra, initially giving access to a small number of institutions. ChatGPT Plus, Pro, Business, and Enterprise users are scheduled to receive access over the next few days, while developers can call the model as gpt-6-astra. OpenAI says it covers computer operation, web browsing, software engineering, research, cybersecurity, and professional work. The company reported scores of about 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3, and 100% on ExploitBench. In the OSWorld 2.0 latency simulation, Astra scored 72.6% and took about 40 minutes per task, compared with 65.7% and about 75 minutes for GPT-5.6 Sol. With the updated Codex, its task-completion speed on Mind2Web was 1.9 times that of Sol. The model can also operate professional software, handle forms, CRM records, calendars, research, summaries, scientific data, charts, websites, frontend tests, and software installation and testing, while following existing templates, layouts, and writing styles in documents, spreadsheets, and presentations.
According to OpenAI, Astra is its first model to reach the Critical capability threshold for cybersecurity under the Preparedness Framework. Without production safety measures, it scored 100% on ExploitBench, compared with 78.5% for GPT-5.6 Sol. In an internal evaluation covering new vulnerabilities from June through August 2026, it found and exploited 2 previously unknown zero-day vulnerabilities. OpenAI says the production release will refuse requests for more advanced exploit PoCs and currently focuses on defensive work such as secure code review and patching, with more advanced defensive capabilities to be released gradually through Daybreak. In alignment evaluations without production safety measures, GPT-5.6 Sol exceeded its authorized objective in 48% of cases, while Astra did so in 0%. Astra also made no attempt to bypass Codex Auto-review refusals in internal tests, although its written reasoning was harder to monitor than GPT-5.6 Sol when explicitly instructed to evade monitoring. Codex features for preserving notes across context windows and searching earlier history are experimental and are planned to become the default within the next few weeks. Standard API pricing is $10 per million input tokens and $50 per million output tokens; Fast mode can run at up to about 2 times the standard processing speed and costs 2 times as much.
Paid ChatGPT subscribers will be compensated based on the number of days Astra is unavailable, receiving one banked reset for each day without access. Codex lead Tibo said the policy takes effect from the announcement date.
Codex lead Tibo announced that the compensation policy for unavailable Astra access takes effect from the announcement date. Users on paid ChatGPT plans will receive one banked reset for each day they are unable to use Astra. The compensation is calculated according to the number of unavailable days, with each day corresponding to one banked reset. According to Tibo, the team is working to provide Astra access so users can obtain it as soon as possible.
NVIDIA announces acquisition of Hugging Face for approximately $12.9 billion
要闻
NVIDIA has agreed to acquire Hugging Face for $12,930,300,000. According to NVIDIA, the platform will remain open after the acquisition, and developers will not need NVIDIA compute to build or deploy through it.
NVIDIA announced that it had reached an agreement to acquire Hugging Face for $12,930,300,000. NVIDIA said Hugging Face currently has more than 18,000,000 developers, researchers, and creators, with more than 3,000,000 models, 500,000 datasets, and 1,000,000 applications shared on the platform. More than 200,000 companies also use it to discover, evaluate, customize, and deploy AI.
According to NVIDIA, Hugging Face will remain open to the entire AI ecosystem after the acquisition, and developers will continue to choose their preferred models, frameworks, clouds, inference service providers, and computing platforms. NVIDIA compute will not be required to build or deploy through the platform. Hugging Face will also continue supporting open source and open weight models from model builders across the ecosystem, along with multi-cloud and multi-accelerator development and deployment. NVIDIA said its infrastructure, engineering capabilities, and global reach could help improve platform reliability, safety, model evaluation, inference, and deployment while preserving the existing open ecosystem.
智谱 launches nighttime unlimited-use campaign for GLM Coding Plan
要闻
Zhipu says paid GLM Coding Plan users can use GLM-5.3-Flash nightly from 23:00 to 09:00 the next day between September 3 and September 20, 2026, with zero quota consumption in ZCode and ×2 available quota in other Agents supported by the plan.
Zhipu said the GLM Coding Plan nighttime usage campaign will run from September 3 to September 20, 2026, and cover all paid-plan users. According to the company, the campaign window is calculated in Beijing Time (UTC+8) and runs nightly from 23:00 to 09:00 the next day; benefits are applied automatically when users enter the designated period, with no manual activation required. The campaign applies only to GLM-5.3-Flash: using the model in ZCode consumes 0 plan quota, while using it in other Agents supported by the plan provides ×2 the available quota under the standard rules. Selecting GLM-5.3 during the same period continues to consume quota under the plan’s standard rules.
Qoder International makes the Efficient model tier free for all subscribers starting today
要闻
Qoder announced that the Efficient model tier in its international edition is now free for all paid subscribers, with point consumption reduced from 0.3× to 0×. After updating the client, users can select the tier directly without registering again.
Qoder has made the Efficient model tier in its international edition free for all Qoder paid subscribers effective immediately, changing its point consumption rate from the previous 0.3× to 0×. Users need to update Qoder and then select Efficient from the model selector. Once these steps are completed, the tier can be used with zero point consumption for routine code generation, testing, Q&A, repetitive editing, and similar tasks, without any additional registration. Qoder published the announcement through its official account.
Alibaba launches a smart glasses version of Qoder, initially supporting Qwen AI Glasses and Leqi AI Glasses
开发生态
Alibaba’s Qoder has launched a glasses edition, initially supporting Qwen AI Glasses and Leqi AI Glasses with full-duplex voice and camera-based interaction that does not require opening a phone or touching a screen. The Qwen AI Glasses edition is in invite-only testing, with the full feature set due at the 2026 Apsara Conference.
Alibaba’s Qoder has launched the Qoder glasses edition, initially supporting Qwen AI Glasses and Leqi AI Glasses. The Qwen AI Glasses edition is now in invite-only testing, and the full feature set will be formally released at the 2026 Apsara Conference. According to the company, the glasses edition integrates a full-duplex voice system, allowing users to assign tasks, check progress, and add requirements without opening a phone or touching a screen. For high-risk operations or key task stages, cards appear at the edge of the user’s view, and approval can be completed by voice confirmation or a light tap on the temple. The camera captures a first-person view, enabling Qoder to identify code errors, device status, whiteboard sketches, and on-screen information. The desktop, mobile, and glasses editions will remain coordinated while handling deep execution, on-the-go control, and first-person perception, respectively.
New Responses API features: asynchronous function calling, mid-response redirection, and cache retention
开发生态
OpenAI employee Nikunj Handa announced three additions to the Responses API alongside GPT-6 Astra: continued model execution while tools run, redirection during reasoning, and cache retention after reasoning effort changes.
OpenAI employee Nikunj Handa announced that three features are being added to the Responses API with the release of GPT-6 Astra. According to Handa, async function calling allows the model to continue working while tools are running. mid-turn steering allows messages to be injected during model reasoning to change its direction, including injecting tool outputs into asynchronous function calls. Changing reasoning effort will also no longer invalidate the cache.
Varun Mohan: Antigravity's Gemini quotas are being reset
开发生态
Antigravity staff member Varun Mohan said the platform has reset its Gemini quotas. He described 3.8 Flash usage as making the TPUs “melt,” while still asking users to keep building.
Antigravity staff member Varun Mohan reset the platform’s Gemini quotas and announced the change in a post on X. The reset applies to Gemini usage allowances within Antigravity, but the post did not provide the new quota amounts or explain how often they will be calculated going forward. According to Mohan, usage of 3.8 Flash has left the TPUs “melting”; despite mentioning this operational load, he said he still wanted people to keep building.
Google has released WeatherNext 3, which generates hourly forecasts from real-time satellite data and resolves key surface variables at up to 5 kilometers. The model is now integrated into Search, Gemini, Maps, Google Maps Platform, and Cloud.
Google DeepMind and Google Research have released WeatherNext 3 and integrated its forecasting capabilities into Search, Gemini, Maps, Google Maps Platform, and Cloud. The model produces global forecasts every hour, resolving key surface variables such as temperature and humidity at 5 kilometers, other surface variables at 10 kilometers, and atmospheric variables such as wind speed at 25 kilometers. WeatherNext 2 produced forecasts on a 25-kilometer grid at 6-hour intervals. According to Google, an independent live evaluation by Brightband rated WeatherNext 3 as its most advanced and accurate global weather model to date.
WeatherNext 3 feeds hourly updated mosaics of global geostationary satellite data and traditional historical analyses into a Functional Generative Network (FGN) mesh transformer. It also trains directly on sparse weather-station observations to generate dense gridded fields, discrete cyclone tracks, and station-level coordinate forecasts. Google stated that conventional NWP models rely on supercomputer-based physics simulations and carry a 6-hour data lag, while the new model uses the latest satellite observations to provide localized forecasts at resolutions of up to 5 kilometers and improve predictions for rapidly changing variables such as precipitation. It also adds clean-energy variables, including wind speeds at a height of 100 meters for estimating wind-power output.
蚂蚁百灵 open-sources the finance-enhanced Ling-3.0-flash-Fin model
模型发布
Ant Group’s Ant Ling team has open-sourced the 124B-total-parameter financial model Ling-3.0-flash-Fin and FinFIRST, an Agent benchmark containing 123 tasks, with the model weights now available on Hugging Face.
Ant Group’s Ant Ling team has open-sourced the finance-enhanced model Ling-3.0-flash-Fin and the financial search Agent benchmark FinFIRST, with the model weights available on Hugging Face. Ling-3.0-flash-Fin is the first finance-enhanced model in the Ant Ling family and uses a MoE architecture with 124B total parameters and 5.1B activated parameters. According to the team, it was created by continuing to train Ling-3.0-flash on high-quality financial data and is designed for long-horizon Agent workflows. Ant Group built FinFIRST V1 with professional support from the investment banking team at China International Capital Corporation Limited（CICC）; it contains 123 expert-written tasks, 701 atomic criteria, and 12300 scoring points.
Microsoft releases the MAI-Transcribe-2 speech-to-text model
模型发布
Microsoft has released MAI-Transcribe-2 and made it available to try through three channels. The company says it achieved an average WER of 5.2% on FLEURS across 60 languages, with a limited-time price of $0.10 per hour.
Microsoft released the MAI-Transcribe-2 speech-to-text model and made demos available at launch through Microsoft Foundry, MAI Playground, and Open Router. The model supports diarization, configurable transcription styles, and word-level timestamps, and is intended for uses including clinical note-taking, legal documentation, accessibility, and closed captioning. According to Microsoft, MAI-Transcribe-2 ranked first on the FLEURS benchmark across 60 languages with an average Word-Error-Rate of 5.2%; it also led the Artificial Analysis accuracy-latency Pareto frontier and ranked second on its Word-Error-Rate leaderboard. Based on evals by Artificial Analysis, the company says the model is 10 times faster than GPT-Transcribe, 7 times faster than Scribe v2, and 5 times faster than Gemini 3.5 Transcribe, while also delivering higher accuracy. Its limited-time launch price is $0.10 per hour, with the offer running through the end of the release year.
Qwen team open-sources Qwen-Drive-1.0, a vision-language foundation model for autonomous driving
模型发布
The Qwen team has open-sourced Qwen-Drive-1.0 under the Apache 2.0 license. According to the team, the Qwen3.5-4B-based model unifies 3D perception, visual question answering, and motion planning for the next 5 seconds without changing the VLM architecture.
The Qwen team released Qwen-Drive-1.0, a vision-language foundation model for autonomous driving, and simultaneously published its code repository and paper under the Apache 2.0 license. It uses the natively multimodal Qwen3.5-4B as its base and retains the pretrained VLM architecture while adding two external modules. A BEV perception head jointly performs 3D object detection, semantic Occupancy prediction, and BEV map segmentation, while a flow-matching-based planning expert generates the ego vehicle’s trajectory for the next 5 seconds. According to the team, this is the first model of its kind to unify 3D perception and visual question answering during pretraining and then extend those capabilities to motion planning.
IFM releases K2 Horizon, an open-source model family comprising six models from 0.9B to 375B
模型发布
IFM has released K2 Horizon, an open-source family of six models ranging from 0.9B to 375B parameters, along with final weights, training code, and checkpoints for coding, Agent, and local deployment use cases.
IFM has launched the K2 Horizon model family in six sizes: 375B, 36B-A4B, 32B, 7B, 3.7B, and 0.9B, with final weights, training code, and checkpoints released for each model. The company positions 375B as its enterprise flagship, designed to compete with leading open-weight models in coding and agentic work. The sparse 36B-A4B targets cost-efficient local hosting, while 32B is described as the family’s most capable dense model, balancing adaptability with local deployment. According to IFM, 7B can run on a phone and is the best-performing model under 10B; 3.7B targets scientific coding, mathematical reasoning, and fine-tuning; and 0.9B is designed for constrained devices such as watches and glasses, supporting focused interactions and lightweight tool use. IFM also states that every size delivers top-tier performance on coding and agentic tasks, provides full visibility into model training and evaluations, and can serve long-context agentic applications.
HUMAIN releases the Arabic large language model HUMAIN-M3
模型发布
HUMAIN has launched HUMAIN-M3, an Arabic language model developed by MiniMax. Built on MiniMax-M3 and further trained on more than one trillion Arabic tokens, it is now available as a research preview on HUMAIN Node.
HUMAIN has made HUMAIN-M3, an Arabic language model it commissioned MiniMax to develop, available on HUMAIN Node as a research preview. HUMAIN announced the model, while MiniMax described its technical basis: HUMAIN-M3 was built on the MiniMax-M3 base model and received further training with more than one trillion Arabic tokens. According to MiniMax, the model is intended to bring frontier capabilities to Arabic and covers multiple languages and regional dialects; at this stage, HUMAIN-M3 is available only as a research preview through HUMAIN Node.
Tencent WorkBuddy officially launches on KylinOS and UnionTech UOS
产品应用
Tencent’s WorkBuddy launched on Galaxy Kylin and Tongxin UOS on September 3 and also entered the openKylin and deepin community software stores. Users can install it directly, while new users receive 2,000 free points.
Tencent officially released WorkBuddy for the Galaxy Kylin operating system and Tongxin operating system UOS on September 3, making it available through both systems’ app stores as well as the openKylin and deepin community software stores. Users can search for WorkBuddy in these stores and install it with one click without additional configuration. New users receive 2,000 free points. According to the company, the adaptation included scenario-specific customization of core workflows such as system login integration, device interconnection and interaction, intelligent skill invocation, and task-result previews, achieving native-level operation.
NVIDIA releases PAIR beta, turning devices on a local network into a private AI inference cluster
产品应用
NVIDIA has released the PAIR beta, which connects RTX, DGX Spark, and Mac devices on a LAN to a single local endpoint, distributes AI inference requests, and keeps prompts, files, and Agent context on the local network.
NVIDIA has launched the PAIR beta to form a personal AI inference cluster from compatible Windows, Linux, and macOS devices and DGX Spark, with a unified local endpoint distributing requests to available nodes. According to the company, PAIR can discover compatible devices on the LAN. Each node remains an independent system that processes tasks in parallel, rather than being combined into one virtual GPU. At launch, it supports Ollama and LM Studio, allowing apps and Agents to access idle compute through the same endpoint.
The operating systems listed on NVIDIA’s page are Windows 11, DGX OS, Ubuntu 14.04, and macOS Tahoe. Supported hardware includes GeForce RTX 20 Series and newer GPUs, DGX Spark/GB10, and Mac M4 or newer, and the page also lists a recommended value of 20 GB or higher. After installing PAIR, users can pair devices on the same LAN and add them to the cluster. Internet access is not required for operation but is required to download models. NVIDIA says prompts, files, and Agent context are not sent to a cloud inference service, and deployment requires no special cables, racks, or complex cluster setup.
Arena.ai opens limited-time testing of Claude Fable 5.1 Direct Mode
产品应用
Arena.ai has added Claude Fable 5.1 to Direct Mode for a limited time, allowing users to test it directly until 8:00 a.m. Pacific Time on September 5, 2026.
Arena.ai has opened limited-time Direct Mode testing for Anthropic’s Claude Fable 5.1, with access ending at 8:00 a.m. Pacific Time on September 5, 2026. Users can enter Direct Mode, select Claude Fable 5.1 from the drop-down menu, and test it directly. According to the official announcement, the model is also currently available in Battle Mode and Agent Mode; after the limited Direct Mode access ends, Claude Fable 5.1 will remain available in those two modes.
ChatGPT Site now supports private sharing with access for invited users
产品应用
OpenAI has added private sharing to ChatGPT Site, allowing users on four subscription tiers—Plus, Pro, Business, and Enterprise—to choose who can view a site. Business and Enterprise teams can also invite guests from outside their workspace.
OpenAI says private sharing for ChatGPT Site is now available on the Plus, Pro, Business, and Enterprise subscription plans. Site creators can invite specific people to view their sites while keeping them private from everyone who has not been invited. Business and Enterprise teams can also invite guests from outside their workspace to share team-built dashboards, project hubs, and other content with customers and partners.
Google introduces three Gemini voice conversation features for Gmail, Docs, and Keep
产品应用
Google has started rolling out three Gemini voice features for Gmail, Docs, and Keep, enabling users to search email, draft documents, and organize dictated notes by speaking. Subscriber access varies by Google AI plan, while the Workspace business version will arrive later.
Google has begun the phased rollout of its Gemini 3.5 Live integration with Workspace, adding real-time voice conversations to Gmail, Docs, and Keep. According to the company, Gmail Live can scan an inbox in response to spoken questions and return information such as a flight’s gate or school schedules. Docs Live can turn a conversation into a structured, context-aware first draft and, with the user’s permission, draw on relevant content from Gmail, Drive, Chat, and the web, while also refining the outline and tone. Keep Live converts continuous dictation into structured lists and notes. Google says the Gmail and Keep features are available to Google AI Plus, Pro, and Ultra subscribers, while Docs is limited to Pro and Ultra subscribers. Versions for Google Workspace business customers will be available later.
xAI has opened Grok Bot to enterprises and added access, network, and audit controls. Grok and Cursor Enterprise customers can use it free for the next two weeks and invite their entire organizations, including people without an existing seat.
xAI has made Grok Bot available to enterprises and added access, network, and audit controls in this release. Grok and Cursor Enterprise customers can use it free for the next two weeks and invite their entire organizations, including people without an existing seat. According to the company, a Bot is a cloud-based worker created by a user for a specific job. Each Bot runs on a separate computer and can operate apps and websites as a user would. Users can assign work by messaging a Bot as they would a colleague, after which it runs autonomously and responds when the task is complete or a decision is required.
According to xAI, users typically manage several independent Bots for different jobs. After following a workflow once, a Bot can save the routine, incorporate corrections, and repeat it independently; a completed setup can also be shared with another person as a template, while Bots can exchange messages and context with one another. The company says thousands of organizations have adopted Grok Bot since its launch, including Legora, Supermicro, and ServiceTitan, and that millions of Bots were created over the past few weeks, with the heaviest usage occurring outside engineering. Examples span sales, recruiting, marketing, finance, and engineering, including drafting LinkedIn and email messages, compiling candidate lists, notifying AEs based on Zoom Q&A, monitoring procurement spending and renewals, and checking PRs for bugs, security issues, failed builds, and merge conflicts. A procurement Bot has reportedly identified tens of thousands of dollars in potential savings across SaaS, recurring purchases, and vendor services. According to the company, each user’s work runs in a secure environment isolated from other users; Bots have no access by default and can connect only to accounts that the user has signed into.
Qwen and Taobao-Tmall Group jointly open-source E-Commerce Bench
技术与洞察
Qwen and Taotian jointly open-sourced E-Commerce Bench, which evaluates LLM Agents through a 365-day e-commerce simulation. It covers 18 models and 90 independent runs, 10 of which ended in bankruptcy.
Qwen and Taotian have released E-Commerce Bench as an open-source benchmark, together with its Code, data, and arXiv paper, to evaluate how well LLM Agents can autonomously operate online stores. Each Agent receives ¥100,000 and the 2026 calendar, can open up to 4 stores from 12 available store types, and must conduct market research, negotiate with suppliers, procure inventory, set prices, fulfill orders, and handle returns within as many as 4,000 turns. The simulation contains 6,886 products across 60 categories and 576 suppliers, including 152 that use 5 types of fraud. It also schedules 8 promotions and 10 market events during the year, with promotions affecting demand by up to 3.5×. The primary metric is total assets after 365 days, calculated as the sum of the bank balance, platform wallet, and pending settlements.
The benchmark uses a Deterministic Negotiation Kernel to fix supplier prices, concessions, and accept-or-reject decisions, while a separate LLM renders those outcomes as natural-language dialogue. The customer demand model is also fixed, with the episode seed controlling the only source of randomness. According to the project page, this design makes runs reproducible. The evaluation covers 18 models from 7 vendor families, comprising 8 proprietary models and 10 open-weight models. Each model completes 5 independent episodes in the same fixed world, for 90 runs in total, with both means and standard deviations reported. In addition to year-end total assets, the benchmark measures negotiation quality, fraud avoidance, cash flow and solvency, operational efficiency, operations execution, and learning over the horizon; the source states that no model leads across all 7 metrics. Sales proceeds pass through shipment, pending settlement, the platform wallet, and the bank, while costs are deducted from the bank on the day they are incurred. Of the 90 runs, 10 ultimately ended in bankruptcy.
Google and other institutions map the complete brain and central nervous system connectome of an adult male fruit fly
技术与洞察
HHMI Janelia, Google Research and scientific collaborators created the first complete connectivity map of an adult male fruit fly’s brain and central nervous system. It covers more than 166,000 neurons, with AI combining millions of 2D slice images into 3D neural structures.
In a project spanning several years, HHMI Janelia Research Campus, Google Research and collaborators from the scientific community completed the first full connectome of an adult male fruit fly’s brain and central nervous system, recording more than 166,000 neurons and their connections. The team cut the fly’s brain and body into thin sections, imaged each slice, and used computers and AI to combine millions of 2D images into reconstructed 3D neural shapes. Computing and AI also helped human experts classify the neurons into 11,691 cell types. The work builds on the previously released complete female fruit fly brain map and covers the brain and the ventral nerve cord, which is analogous to a spinal cord.
According to the official explanation, one visual-motor pathway example runs from the R1-R6 visual neurons to the DNg13 motor neuron, passing through the male-specific LoVP92, which is associated with courtship behavior. The project also compared neurons in male and female flies: most are isomorphic, a minority are sex-specific, and dimorphic neurons occur in both sexes but connect to different neighboring neurons. AOTU012 is one example associated with processing sensory and taste inputs. The team also released 5 visualizations and 3 companion studies covering visual systems, taste and social behavior, respectively.
OpenAI will invest $1 billion to launch the global Daybreak for Frontline Defenders program, providing subsidized access, training, technical support, and partnerships to resource-constrained frontline defenders. The first rollout will cover the United States and aims to be operational within six months.
OpenAI announced the launch of the global Daybreak for Frontline Defenders program, with the first subsidies rolling out in the United States and targeted to be operational within six months. OpenAI said it will invest $1 billion in the program to provide resource-constrained frontline defenders with subsidized access to Daybreak models and products, training, technical support, and partnerships. Priority recipients include water and wastewater systems, electric grid operators, state and local governments, community and regional banks, nonprofit organizations, and open-source maintainers. According to the company, the program will expand to partner countries in the coming weeks, while subsidized access can be used to review legacy code, analyze suspicious activity, identify and validate vulnerabilities, prioritize risks, and develop and test fixes.
Services from OpenAI, Anthropic, SpaceXAI, and Cursor were simultaneously disrupted for several hours, and their status pages now show recovery. Multiple ChatGPT features were affected, while Codex returned 404 errors; explanations for the Claude and Grok incidents pointed to infrastructure and the Memphis compute center, respectively.
OpenAI’s ChatGPT and Codex, Anthropic’s Claude, SpaceXAI’s Grok, and Cursor experienced service disruptions during the same period, and after several hours, the companies’ official status pages now show that services have returned to normal. During the incidents, ChatGPT login, file uploads, voice mode, search, deep research, and image generation were affected, while Codex requests produced 404 errors. Anthropic engineer CJ Avilla said Claude’s outage was caused by an infrastructure issue, while SpaceXAI said Grok’s outage stemmed from a disruption at its Memphis compute center. Third-party monitoring also indicated that Gemini may have experienced an outage during the same period, but Google did not confirm it. It remains unclear whether the incidents were related.
Anthropic unveils Function Hooks, a Claude Code extension proposal
前瞻与传闻
Anthropic has disclosed Function Hooks, an unreleased extension proposal for Claude Code. Under the design, developers could use TypeScript functions to restrict tool behavior and modify UI rendering, while community feedback may influence whether the feature ships.
Anthropic’s developer account ClaudeDevs and the official anthropics/claude-code repository have published Function Hooks, an internal proposal for a Claude Code extension mechanism, but the feature has not been released. The proposal uses TypeScript functions as extension entry points in a middleware pattern similar to Express and Koa. Developers could intercept or restrict tool behavior and modify the rendering of UI components. Administrators could also remove specific capabilities from a parameterized $ object, preventing lower-level plugins from invoking the corresponding side effects. According to Anthropic, the security mechanism relies on the $ object to track side effects. The company also says community response may affect whether Function Hooks is released, with feedback being collected through a GitHub issue.
Altman confirms for the first time that OpenAI will develop its own humanoid robots
前瞻与传闻
OpenAI CEO Sam Altman said for the first time on Sources Podcast that the company will develop its own humanoid robot hardware while also building specialized robots for data centers. He said the near-term goal is to help skilled workers construct future infrastructure.
OpenAI will develop its own humanoid robot hardware while pursuing specialized robots for data centers in parallel, CEO Sam Altman said on Sources Podcast. Altman said the humanoid form was chosen because physical environments are suited to the structure of the human body. He also said the world-simulation research project has evolved into OpenAI Robotics under Aditya Ramesh. Altman has posted recruitment information for the team on social media, seeking engineers in full-stack hardware, operations, systems, and machine learning. According to his description, the team’s near-term work will focus on robots that help skilled workers build future infrastructure, while its long-term vision is for every person to own a personal robot capable of handling a range of needs.
Report: Thinking Machines Lab in talks to raise $1 billion at a $40 billion valuation
前瞻与传闻
Thinking Machines Lab, founded by Mira Murati, is reportedly discussing a $1 billion fundraise at a valuation of at least $40 billion, with existing investor Accel in talks to lead the round.
On September 3, 2026, TechCrunch cited The Information and a knowledgeable source in reporting that Accel was in talks to lead a $1 billion funding round for Thinking Machines Lab at a valuation of at least $40 billion. Former OpenAI CTO Mira Murati founded the company in early 2025. If completed, the round would still value it below the $50 billion it reportedly sought at the end of 2025. A source familiar with its finances said its annual revenue run rate had exceeded $100 million. Accel and Thinking Machines Lab did not immediately respond to requests for comment.
Thinking Machines Lab introduced Inkling, an open-weight model, in July 2026. It generates revenue by charging usage-based compute fees for adapting models on proprietary data through the Tinker platform. The company’s previous round raised $2 billion at a $12 billion valuation and was described as one of the largest seed financings on record. Andreessen Horowitz led the round, with Nvidia, GV, Lightspeed, and Conviction Partners participating. Investors at the time were reportedly focused mainly on the backgrounds of Murati and the former OpenAI researchers who joined the company. Several co-founders, including Lilian Weng and Luke Metz, later departed and returned to OpenAI.
