Google has opened Gemini 4 Carbon to paying developers and enterprises and is rolling it out to Google AI Ultra subscribers. Its introductory API pricing matches Gemini 4 Argon at $2 per million input tokens and $10 per million output tokens.
Google has made Gemini 4 Carbon available to all paying developers and enterprises through the API in Google AI Studio and Gemini for business, and has begun rolling it out to all Google AI Ultra subscribers in Google Antigravity, the Gemini app, and AI Mode in Search. The company positions Carbon for complex, multi-step coding, debugging, and open-ended scientific research, saying it uses more tokens than Gemini 4 Argon for exploration, reasoning, and verification. It continues to position Argon as the suitable choice for most software engineering and knowledge work.
According to Google, Carbon ranks first on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench, with a per-task cost less than half that of the next-best external model. On the Multimodal Knowledge Agent (MKA) benchmark launched with Artificial Analysis, Carbon ranks second, with an average per-task cost less than one-sixth that of the top-ranked model. Google is also previewing /swarm in Antigravity, allowing a single prompt to drive a multi-Agent team that can be tracked in real time. Carbon can handle architecture design, coordination, and auditing alongside faster, lighter models working in parallel isolated workspaces, with adversarial verifiers checking each change.
MiniMax has released H3-Image, an H3-based image generation model, and made its weights available, with support for up to 16 reference images. The company says it generates 4K images in under 10 seconds and costs 50–70% of comparable mainstream services.
MiniMax has released the H3-based H3-Image and made its model weights available; the source does not specify a release date. The model supports text-to-image generation, reference-based generation, and image editing, and is also available through MiniMax API Platform, MiniMax Code, and Hailuo AI. According to the company, it can process up to 16 reference images and generate a 4K (16MP, 4096 × 4096) image in under 10 seconds, at up to 5× the speed of mainstream alternatives.
According to MiniMax, H3-Image’s optimizations span the entire post-training pipeline, including data, reinforcement learning, sampling-step reduction, PE, and deployment optimization. The company says Standard mode runs at 1.5–2× the speed of mainstream alternatives and Highspeed mode at 3–4×, with pricing at 50–70% of comparable mainstream services. During launch week, image generation requests at 4K resolution and above are charged at 50% of the standard price; the source does not specify the promotion’s start and end dates or absolute prices.
Anthropic introduces Claude Whiteboards, Claude Sheets, and Claude Data
要闻
Anthropic introduced Claude Sheets, Claude Whiteboards, and Claude Data and upgraded interactive connectors to let partners embed apps in Claude. Its Marketplace lists more than 9,000 connectors and plugins.
Anthropic launched Claude Sheets, Claude Whiteboards, and Claude Data, making new artifact types and an interactive connectors upgrade available with the announcement; the source does not specify a publication date. According to the company, teams can build spreadsheets, whiteboards, and dashboards with Claude, as well as custom account trackers, daily briefings, or contract review tables. It says the same context can carry across different artifacts without users having to repeat the details.
For app integrations, the company says partners can now embed apps directly in Claude, allowing users to work in apps such as Adobe and Atlassian without switching tabs. Claude accesses external context through connectors built on the Model Context Protocol (MCP). Its Marketplace lists more than 9,000 connectors and plugins, including community-built offerings. Plugins from Salesforce (beta), NetSuite, and Workday (beta) include built-in skills. Claude Meetings is not yet available; the company says it will support taking notes during calls and updating documents that teams are working on together.
Anthropic offers a free full usage reset after excess consumption of compensation credits
要闻
Anthropic has provided all Pro, Max, Ultra, and Team users with one free full usage reset covering every model and usage limit. The deadline to claim it is December 10 at 14:00 Pacific Time, and the restored allowance also works in Claude Code.
Anthropic has issued one free full usage reset to all Pro, Max, Ultra, and Team users, to be used before December 10 at 14:00 Pacific Time; the source does not specify a year. The reset covers all models and usage limits. According to the company, it compensates for several fixed issues that caused higher-than-expected usage consumption, and the earlier Fable reset did not sufficiently compensate users who had exhausted their weekly allowances across all models. The Reset for free button is available on the Usage page under Settings in Claude’s web, desktop, and mobile apps. The restored allowance also applies to Claude Code, but its terminal and IDE interfaces currently have no button to claim the reset.
OpenAI enables third-party tools to use ChatGPT Business subscriptions
要闻
OpenAI now lets users access their ChatGPT Business subscription in third-party tools, including Notion and v0, through Sign in with ChatGPT. According to its official guidance, usage counts toward existing limits, and connecting an app does not add any allowance.
OpenAI announced that ChatGPT Business subscriptions are now available for use in third-party tools through Sign in with ChatGPT, with support for Notion, v0, ampcode, kilocode, Warp, Nous Research’s Hermes, openclaw, and opencode. Workspace administrators can enable access for individual tools in the admin console and set usage limits as a percentage of each member’s included allowance. According to the official documentation, in-app usage draws from the subscription’s existing limits, and connecting an app provides no additional allowance. Using the subscription also does not grant apps access to users’ ChatGPT conversations or memories.
Google DeepMind has made Gemini Robotics 2.5 available to Google’s robotics partners; according to the company, the model achieved a 96% success rate in a proof of concept involving metal-part sorting on automotive production lines.
Google DeepMind has introduced Gemini Robotics 2.5, a vision-language-action (VLA) robotics model, which is now available to Google’s robotics partners. The company says the model improves robot dexterity, reliability and safety. In a proof of concept conducted with Apptronik and Mercedes-Benz, Apollo 2 humanoid robots used the model to sort metal parts on operating automotive production lines in Hungary and Germany, achieving a 96% success rate. Google DeepMind tested the model on more than 200 tasks; according to its evaluation, success rates were 30% to 100% higher than the previous baseline, depending on task type, while task completion was 20% to 30% faster. The model also incorporates safety mechanisms including hazard recognition and jailbreak prevention; according to the company’s evaluation, safe task completion rates improved by 33% around obstacles and 43% around people.
ElevenLabs releases Scribe v3 and Scribe v3 Live transcription models
模型发布
ElevenLabs has introduced Scribe v3 and Scribe v3 Live, both available through ElevenAPI. In company evaluations against their respective predecessors, Scribe v3 reduced WER by 27%, while Scribe v3 Live reduced semantic WER by 43%.
ElevenLabs has released Scribe v3 and Scribe v3 Live, with both available through ElevenAPI at launch and the Live version also available in ElevenAgents; the source does not specify a release date. According to the company, migrating from Scribe v2 requires changing just one line containing the model ID. For output control, Scribe v3 lets users choose whether numbers appear as digits or words, accepts up to 1,000 terms through keyterm prompting, and supports transcript formatting through natural-language rules. The Live version supports language switching within a sentence and uses conversational context alongside audio to determine when a speaker has finished. In evaluations across 102 languages on an internal version of FLEURS, the company reports that Scribe v3 reduced WER by 27% compared with Scribe v2; Live reduced semantic WER by 43% and WER by 31% compared with Scribe v2 Realtime. These results average scores calculated separately for each language, giving all 102 languages equal weight.
Amazon open-sources tabular foundation model Mitra-v2
模型发布
Amazon has open-sourced Mitra-v2, a tabular foundation model supporting classification and regression, along with inference and fine-tuning code under Apache-2.0. According to the company, the model was pretrained exclusively on synthetic data and ranks first among single models on TabArena, which covers 51 real-world datasets.
Amazon has released its second-generation tabular foundation model, Mitra-v2, making both classification and regression versions, along with inference and fine-tuning code, available on Hugging Face under Apache-2.0. According to the official description, pretraining uses only synthetic data, with no real-world or customer data. By default, the model automatically performs brief fine-tuning on each new table, generates 8 copies, and averages their predictions. On performance, the company says Mitra-v2 ranks first among single models on the public TabArena benchmark, which contains 51 real-world datasets. It also reports that Mitra-v2 outperforms Mitra-v1 on 32 of the 33 datasets that both generations can run on natively, while the fine-tuned version ranks first on the ScoringBench regression benchmark.
Google Antigravity previews /swarm for parallel development with collaborating agents
开发生态
Google Antigravity has introduced a preview of /swarm, allowing multiple Agents to divide tasks, work in parallel and coordinate through a shared message board. Available in the App and CLI, it uses a coordinator to assign work and report results; the source gives no launch date.
Google Antigravity has introduced a preview of /swarm for its App and CLI, without specifying a launch date in the source. According to the official documentation, running /swarm turns the current conversation into the coordinator, which assigns tasks, evaluates returned results and reports to the user rather than performing technical work itself. The coordinator determines the number of workers, the files each owns and when to bring in a verifier. Members coordinate through a shared message board, while users can continue checking progress or changing instructions in the main conversation. The documentation identifies tasks that can be split independently across files, packages or questions as suitable uses.
App users can open the Swarm pane through a link in the coordinator’s reply. CLI users can run /swarm board or press Alt+S to view the live board, which displays member states, their models and posts. The coordinator also adds @monitor, which posts only when work remains unfinished and all members have gone silent, calling on relevant members to move the work forward. A member pauses when it exhausts its quota and does not automatically restart when the quota refreshes; a new board post, a model switch or a direct message to that member triggers resumption. According to the documentation, waiting workers consume no tokens, and removing a worker is permanent.
Meituan's LongCat 2.5 is free on OpenCode for two weeks
开发生态
OpenCode announced a two-week period of free access to Meituan’s LongCat 2.5, citing support for a 1M context, text and image inputs, and zero data retention. The platform’s data page ranked the model 26th by usage last week.
OpenCode announced that Meituan’s LongCat 2.5 is now available free on its platform for a two-week period. The announcement lists a 1M context, multimodal input, and zero data retention among the model’s supported capabilities; zero data retention is a statement published by the platform. Input types and usage figures appear on OpenCode’s data page, which lists text and images as supported inputs and places the model 26th by usage on OpenCode last week.
Tailscale OAuth apps become generally available, enabling Grok Bot to access private-network MCP servers
开发生态
Tailscale has made OAuth apps generally available, enabling Grok Bot to access MCP servers on users’ private networks through OAuth 2.0 authorization without exposing those servers to the public internet. Tailnet policies continue to control the scope of access.
Tailscale has launched OAuth apps into general availability, and Grok Bot is already using the mechanism in production to connect to MCP servers within users’ tailnets. The source does not provide a specific release date. According to the company, users can connect through the Tailscale plugin in the Grok Bot desktop app without exposing internal servers to the public internet or pasting keys or other secrets. OAuth apps use the standard OAuth 2.0 authorization flow to provision devices under the authorizing user’s identity, preserving tailnet access policies, audit logs, and device quotas. Existing OAuth clients, by contrast, create tag-owned devices under a shared service identity.
Access can be restricted on a per-device basis without giving Grok Bot all the network permissions available to a user’s other devices. According to the company, Premium and Enterprise plans provide Grok Bot devices with a custom posture attribute, allowing access to resources to be controlled through device posture checks. On other plans, users can assign a tag to the device after authentication and use tag-based access rules. The company’s example uses defaultSrcPosture to prevent existing access rules from applying to Grok Bot, then uses an explicit srcPosture to allow access to designated MCP servers. Resources for developers building their own integrations include the OAuth Apps documentation and device provisioning guide.
Chrome joins Firefox, Edge, and MDN to release Modern Web Guidance 1.0
开发生态
Chrome released Modern Web Guidance 1.0, with Firefox and Edge joining the project to bring browser compatibility knowledge into coding agents and add a modernization workflow for existing code.
Chrome released the stable version of Modern Web Guidance 1.0 on November 10, 2026, with Firefox and Edge joining the project to provide support, improve compatibility, and contribute. The tool supplies coding agents with web platform knowledge through a set of skills, with Baseline and browser compatibility data embedded in its guides. According to the official announcement, it can recommend features based on a project's browser support requirements, and its recommendations are maintained by browser engineers, web standards experts, and accessibility specialists. The modernization workflow added in 1.0 can analyze existing HTML, CSS, JavaScript, and package.json files to identify upgrade opportunities and guide an Agent through changes; the initial release only uses static analysis to check a limited set of known patterns, while agentic audits that reason directly about code are planned for later. The announcement also states that the user-facing guides in 1.0 were reviewed and tested with real assistive technologies by accessibility experts who rely on them daily.
OpenAI dots team member pash announced that Dots has launched multi-exec, allowing a dot to run commands on a connected computer and use its resources and tools, with spawned subagents sharing the same computer.
According to a post on X by OpenAI dots team member pash, Dots has launched multi-exec, enabling a dot to execute commands directly on a connected computer. According to pash, the dot can reuse that computer’s compute resources, storage, logged-in browser sessions, CLI tools, and locally saved skills, while subagents spawned by the dot share the same computer. OpenAI had previously announced that a single dot could connect to multiple local computers and select which one to use based on the task.
OpenAI has launched the Businesses in ChatGPT beta, allowing companies to maintain business and product information for ChatGPT’s 1.2 billion weekly users, while making Business Agent available to eligible US businesses.
OpenAI has launched the Businesses in ChatGPT beta, enabling companies to claim their organizations through Business Hub, update key information after verification, and add product data sources containing prices and stock availability themselves. Hub’s aggregated insights cover business mention counts, website clicks, and common themes in conversations that reference a business, but are provided only when relevant interactions reach a certain threshold. Eligible US e-commerce and online service businesses can also create a Business Agent; when a business appears in an answer, users can open its profile and ask questions in a clearly labeled, separate conversation. OpenAI says it will not share users’ conversations or the contents of Business Agent conversations with businesses.
Meta Muse gets its own email address to send and receive messages on users' behalf
产品应用
Meta’s personal AI agent Muse now offers a dedicated email address to paid subscribers in the United States and Canada, allowing it to send and receive email on their behalf. Users can hand off follow-up handling by adding it to an email thread or forwarding messages to it.
Meta’s personal AI agent Muse has introduced a dedicated email address, with email sending and receiving on users’ behalf now available to paid subscribers in the United States and Canada. According to its explanation, users can start using the email feature by opening the mail tab through their profile picture, or by asking Muse directly. Messages that users want Muse to handle can be forwarded to it; users can also add Muse to an existing email thread to have it help take over subsequent handling.
Adobe previews local MCP integration, enabling control by third-party agents
产品应用
Adobe announced a local MCP at MAX to let third-party chat tools access more Creative Cloud functions, initially spanning four apps. The company is still determining the scope of access and will adjust it based on user feedback.
As reported on November 10, 2026, Adobe announced a local Model Context Protocol (MCP) at MAX in Miami to expose more Creative Cloud functions to third-party chat tools. The announced endpoints and capabilities span Illustrator, Photoshop, After Effects, and Premiere. Adobe already provides access to some product functions through ChatGPT and Claude integrations. Adobe vice president Deepa Subramaniam said the company is still determining which capabilities to expose and where to draw boundaries, and will observe users’ approaches and goals while opening up further functions based on feedback.
Other updates announced at the event include Builder, a prototyping tool in private beta in Firefly, alongside tools for creating on-brand graphs, editing selected parts of images, and assembling videos. According to Adobe, the video tool can combine clips, images, and sound effects, and the company also plans to keep its Agent running continuously and proactively gathering work context. At the application level, Photoshop will gain relighting and sliders for prompt-based generative styles; Premiere will add a professional color grading system and video object removal; and Lightroom will add tools for correcting closed eyes and removing blemishes.
Adobe and Google expand partnership with two-way integration of tools and AI models
产品应用
Google announced an expanded partnership with Adobe at Adobe MAX, with plans to bring Adobe’s creative and productivity tools into Google products and Google’s AI models into Adobe apps through a two-way integration.
Google announced at Adobe MAX that it is expanding its partnership with Adobe, with the companies planning a two-way integration of tools and AI models, but the announcement provided no specific launch date. According to the official announcement, Adobe’s creative and productivity tools will be incorporated into Google products used in everyday workflows, while Google’s frontier AI models will be brought into Adobe apps. The announcement identifies switching apps to edit PDFs or fine-tune images as use cases the partnership aims to address, and states that switching apps interrupts users’ focus. It does not name the specific products or models to be integrated.
Google Cloud launches Gemini for healthcare in US preview
产品应用
Google Cloud has announced the U.S. preview of Gemini for healthcare, bringing medical records, insurance policies, and administrative workflows into a shared Agent workspace. The company says the solution supports HIPAA compliance while preserving existing access permissions.
Google Cloud announced Gemini for healthcare and opened its U.S. preview; the source does not specify the announcement date. The solution extends the general-purpose Gemini agent for healthcare providers and health insurance organizations. According to the company, it can carry out multi-step tasks across clinical, financial, and administrative systems under human oversight, including organizing patient histories, preparing visit materials, assembling authorization packets against payer policies, and checking requests against medical policies for reviewers. Its components include reusable Skills, system Connectors, a partner ecosystem, and a unified governance platform, with organizations able to customize Skills to their own policies and operating procedures. The company says Connectors link EHR, claims, and management systems without requiring sensitive data migration or changes to existing permissions. The platform supports HIPAA compliance, provides end-to-end audit logs and verifiable citations, and allows organizations to select Google and partner models under the same governance framework. For insurance services, it can also use plan documents to estimate out-of-pocket costs, track deductibles, and verify in-network eligibility.
Google DeepMind uses AI to verify core Signal protocol code in four weeks
技术与洞察
Google DeepMind released AlphaProof Code, completing machine-code-level formal verification of the core cryptographic code in Signal’s reference implementation in about four weeks. The system generated approximately 6.5 million lines of Lean proof code, and the verification found no new vulnerabilities.
Google DeepMind released the AI framework AlphaProof Code and completed machine-code-level formal verification of the core cryptographic code in libsignal, the Signal protocol’s reference implementation, in about four weeks; the source provides no release date. The verification did not cover the entire Signal application and found no new vulnerabilities. Humans and AI jointly wrote mathematical specifications, after which agents powered by Gemini Carbon generated approximately 6.5 million lines of Lean proof code for automated checking by Lean. Only 1% required human review. The team also reintroduced two previously fixed vulnerabilities into isolated versions, and in both cases the system produced machine-checked results proving those versions unsafe. According to the official account, this is the first method to connect protocol-level security guarantees to the unmodified code of a widely deployed protocol.
Google DeepMind reveals two AI cryptanalysis advances
技术与洞察
Google DeepMind disclosed two cryptanalysis results produced by AI agents under human guidance, identifying two less computationally expensive attack paths against legacy 1024-bit RSA and reducing CPU computation for a signature experiment by approximately 72%.
Google DeepMind announced two AI-assisted cryptanalysis research results in which human researchers guided agents to explore attacks against legacy 1024-bit RSA. The team's proposed Fan Factories algorithm reduces the theoretical cost of each subsequent 1024-bit target in batch factorization to approximately 1/30 of the previous method's cost. The other study, focused on signatures, found that padding protection fails when an attacker can control approximately half or more of a signature data block. Where less content is controllable, the new algorithm's theoretical computational cost is approximately 1/1600 of the previous method's. Two agents reproduced the signature experiment in less than 4 days, whereas the research team had previously needed approximately 5 months to complete the same experiment. CPU computation fell from approximately 1380 core-years to 391 core-years, a reduction of approximately 72%. According to the official statement, these results affect only obsolete 1024-bit systems, while modern encryption protocols using keys of 2048 bits or more remain secure.
Google and Artificial Analysis introduce the MKA benchmark
技术与洞察
Google and Artificial Analysis have jointly launched MKA, a benchmark with 200 tasks assessing how Agents handle visually complex files. It covers 10 industry verticals and 83 occupations, with a separate set of 10 public tasks available for testing.
Google and Artificial Analysis have jointly released the Multimodal Knowledge Agent (MKA) benchmark, with model evaluation available on Artificial Analysis. According to the official announcement, it targets knowledge work requiring an understanding of visual information such as images and page layouts. Experts across 10 industry verticals and 83 occupations helped design its 200 tasks. Each task represents a multistep project that takes an experienced professional three to eight hours to complete and includes input files and a rubric of roughly 70 yes/no checks, weighted 75% toward factual accuracy and 25% toward visual design. All models use the open-source Agent harness Stirrup, with identical tools, sandbox, and turn budgets.
An Agent-based autorater built on a non-Google model handles scoring, examining files, rendered pages, and spreadsheet formulas. According to the announcement, domain experts and the autorater agreed on grading 96% of the time. Artificial Analysis keeps the 200 evaluation tasks private to prevent their inclusion in training data or discovery by models through search; a separate set of 10 public tasks is available on Hugging Face. The official report states that tested models used an average of 30 to 130 turns per task. Gemini 4 Carbon ranked 2nd in quality, with an average task cost one-sixth that of the highest-scoring model. It is available through the Gemini API in Google AI Studio.
Anthropic kicks off IPO roadshow, plans to offer 825 million shares
行业动态
Anthropic has launched its IPO roadshow, proposing to offer 825 million shares of Class A common stock at an expected price of $70 to $78 per share. It has applied to list on the Nasdaq Global Select Market under the ticker ANT.
Anthropic announced the launch of its initial public offering (IPO) roadshow, with the registration statement for the securities filed with the U.S. Securities and Exchange Commission but not yet effective. Under the company's announced plans, the offering would comprise 825 million shares of Class A common stock at an expected price of $70 to $78 per share. Anthropic plans to offer approximately 801 million shares, with certain existing shareholders offering the remainder. The company has applied to list on the Nasdaq Global Select Market under the proposed ticker ANT. The securities cannot be sold, nor can offers to buy be accepted, before the registration statement becomes effective.
Stripe announced that OpenRouter has officially joined the company; OpenRouter says its service is used by more than 10 million developers and companies. According to Stripe’s CEO, OpenRouter’s weekly token volume has grown by approximately 125% since the acquisition announcement, and the companies plan to integrate capabilities over the coming months.
Stripe announced that it has completed its acquisition of OpenRouter, bringing OpenRouter officially into the company. According to OpenRouter, more than 10 million developers and companies currently use its service, and it will maintain its neutrality after the acquisition, with models and providers competing on quality, reliability and price. Stripe says it is building infrastructure for the AI economy and will work with OpenRouter to help businesses explore new models and optimize token routing and usage. Stripe CEO Patrick Collison said OpenRouter’s weekly token volume has grown by approximately 125% since the acquisition was first announced, and the companies plan to integrate their capabilities over the coming months.
OpenAI has acquired travel AI Agent company Odessia, which says its service is no longer available. Users can still contact support@odessia.com for help changing or modifying existing bookings or accessing existing travel plans.
OpenAI has acquired travel AI Agent company Odessia, and Odessia has discontinued its service as part of the transition. In an announcement on its website, Odessia said it was founded to reinvent the travel experience and described the acquisition as the next stage of that mission. The company continues to offer a support channel for users with existing bookings or travel plans. According to its announcement, users who need to change or modify existing bookings, or access existing travel plans, can contact support@odessia.com for assistance.
