Google has launched Gemini 4 Hydrogen for everyday coding and Agent workflows across access points including the Gemini API. Pricing is flat across its 1 million-token context window, at $1.50 per million input tokens and $7.50 per million output tokens.
Google introduced Gemini 4 Hydrogen in an announcement that did not specify a publication date, and announced Orchestrator in Google Antigravity to dynamically switch between models such as Carbon and Hydrogen within the same session. Hydrogen’s availability includes all tiers of Google AI Studio, the Gemini API, Gemini for business, the Gemini app, Google Search, and paid Google AI Pro and Ultra subscribers in Google Antigravity. Pricing is flat across its 1 million-token context window, at $1.50 per million input tokens and $7.50 per million output tokens.
According to Google, Hydrogen offers four thinking levels—low, medium, high, and xhigh—with high as the default. On the Artificial Analysis Intelligence Index, Hydrogen scores 51 at high, compared with Argon’s 50, and reaches 54 at xhigh. Google says Orchestrator achieved task success rates close to Carbon’s on Terminal-Bench 4.0 while consuming 24% less quota on average and completing tasks 19% faster. For parallel Agent workflows, Carbon can coordinate while Hydrogen serves as worker and verifier agents, building and testing changes simultaneously in isolated workspaces.
Google Antigravity introduces an Orchestrator option
要闻
Google is introducing Gemini 4 Hydrogen and Orchestrator in Antigravity. According to the company, Orchestrator achieves near-Carbon task success while using 24% less quota on average and completing tasks 19% faster, with available options determined by subscription plan.
Google announced in a post without a specific date that it was beginning to roll out Gemini 4 Hydrogen and an Orchestrator option in the Google Antigravity model selector, with access determined by subscription plan. Paid Google AI Pro and Ultra users can access Hydrogen. Google AI Ultra users receive Orchestrator (Carbon) and Orchestrator (Argon), while paid Google AI Pro users receive Orchestrator (Argon). Enterprises can access Hydrogen and Orchestrator through Gemini for business.
According to Google, rather than assigning a fixed model to an entire task based only on the initial prompt, Orchestrator supports bidirectional handoffs between Carbon and Hydrogen as the same task progresses. For example, Carbon can first examine dependencies, resolve ambiguities, and develop a plan in a read-only phase, before handing off to Hydrogen to edit code, update packages, and run tests. If Hydrogen encounters obstacles, it can hand control back to Carbon to inspect the state and revise the plan. Google says that on Terminal-Bench 4.1, Orchestrator achieved near-Carbon task success rates while consuming 24% less quota on average and completing tasks 19% faster.
NVIDIA’s board announced Jensen Huang’s immediate resignation as President, CEO and a director. CFO Colette Kress will also serve as Interim CEO while the board seeks a permanent successor from inside and outside the company and appoints its first board chair.
NVIDIA’s board announced that Jensen Huang’s resignation as President, CEO and a director takes effect immediately, with Executive Vice President and CFO Colette Kress becoming Interim CEO. Kress will retain her existing positions until a permanent successor is appointed; the board has begun the search and will consider internal and external candidates. The board also appointed Harvey C. Jones as the company’s first Chair of the Board. Jones previously served as CEO and Executive Chairman of Synopsys and has more than 30 years of experience in entrepreneurship, technology company leadership and venture investing. According to the board, the company expects its operations, customer commitments and product roadmap to remain unaffected by the leadership changes. Kress said the company would continue its work in accelerated computing and AI, serving customers, partners, developers and shareholders.
ChatGPT dot update adds support for custom instructions and behavior rules
要闻
Official ChatGPT dot documentation describes separate custom instructions and optional behavioral rules for setting boundaries on execution, approvals, and handoffs. Rules cannot expand app access or bypass built-in safety requirements, and the documentation gives no effective date.
Official ChatGPT dot documentation describes controls using custom instructions and optional custom rules, without specifying an effective date. According to the documentation, custom instructions apply only to the relevant dot and are separate from ChatGPT custom instructions. Users can set or update preferences through conversation, including writing style, reporting frequency, and when to ask before acting. Custom rules govern specific actions, allowing users to authorize execution, require prior approval, or hand an action back to the user. Saved rules apply to the dot within the same account, but do not grant access to apps or computers, override built-in safety requirements, or remove required confirmations such as approval to use a saved login. The documentation also states that a dot may fail to follow rules correctly. If a workspace disables custom rules, saved rules no longer apply and cannot be edited.
According to the documentation, actions affecting accounts or sharing information undergo an automatic review against instructions, permissions, rules, and safety requirements to determine whether they can proceed, need approval, or require the user to act. Requesting a draft reply does not authorize sending it. On desktop, Activity shows task progress, files, and results, while Scheduled shows scheduled tasks’ instructions, timing, and destinations. A dot can research information from connected apps, but its research tools cannot send messages, modify app content, or control a browser or computer. OpenAI says it does not train models directly on proactive research or its private notes; if that information enters an eligible conversation or task, the user’s data settings apply. Stopping a task does not undo completed actions, and deleting a dot does not reverse changes made in apps or recall messages already sent.
Tibo, a member of OpenAI’s Codex and ChatGPT teams, confirmed that a banked reset is active and will roll out to all user accounts. Users have posted screenshots saying they have received this usage reset.
Tibo, a member of OpenAI’s Codex and ChatGPT teams, confirmed in a follow-up post the day after his teaser that the banked reset was active. According to his explanation, the usage reset would subsequently roll out to all user accounts. Users have also reported receiving it, with screenshots accompanying posts stating that the reset had arrived. In his earlier teaser, Tibo said it had been a long time since a usage reset had been issued. He coupled a remark about perhaps needing to leave his basement with a suggestion that Thursday might be a suitable day for a banked reset.
ByteDance’s Volcano Engine has released Seedance 2.5 flash, supporting videos up to 30 seconds long. Text-to-video and image-to-video generation at 720P cost approximately 0.99 yuan per second, while the company says generation is twice as fast as Seedance 2.5.
ByteDance’s Volcano Engine has officially released Seedance 2.5 flash, although the source does not specify the release date and lists November 15, without a year, as the launch date for its API on Volcano Ark. The model offers 480P and 720P resolutions, supports videos up to 30 seconds long, and accepts up to 10 images, 5 video clips and 5 audio clips per input. Text-to-video and image-to-video generation at 720P cost approximately 0.99 yuan per second, approximately 34% less than Seedance 2.5. According to the company, its overall output quality exceeds that of the previous flagship model, Seedance 2.0, and is the closest to Seedance 2.5 among current models, while generation is twice as fast as Seedance 2.5. For platform availability, the company says rollout to Jimeng, LibTV, Updream, Creao AI, TapNow and other platforms will begin on the release date.
豆包工作 integrates multiple third-party models and launches a limited-time promotional plan
要闻
Doubao Work has announced support for third-party models including Kimi K3.1 Pro, GLM 5.3 and DeepSeek V4.1 Flash, alongside limited-time quota discounts for two models. New users can also claim 30 days of subscription benefits for free by logging in to the desktop version.
Doubao Work has announced a product update adding several third-party models alongside its in-house Doubao 2.1 Lite and Pro models. The source does not specify the update date. The additions include Kimi K3.1 Pro, GLM 5.3 and DeepSeek V4.1 Flash. According to the company, the model selection page will display each model’s quota consumption multiplier, with lower multipliers indicating lower quota consumption.
On quota usage, the company says it will introduce a limited-time promotion for participating models during designated hours on weekdays and holidays. Limited-time discounts are already available for Doubao 2.1 Lite and DeepSeek V4.1 Flash, with reduced quota consumption multipliers. According to the company, this reduces the quota required for the same task. The source does not list the specific discount hours, multipliers or promotion end date. New users can also currently claim 30 days of subscription benefits for free by logging in to the desktop version of Doubao Work.
Anthropic has begun rolling out Claude Meetings beta to Team and Enterprise plans, with support currently available for Google Meet. Claude can participate in calls and help carry out tasks, while meeting notes and transcripts are free during the beta.
Anthropic has started rolling out Claude Meetings beta on Team and Enterprise plans. Google Meet is currently supported, with support for other video conferencing platforms still to come. According to the company, Claude joins as a visible participant and posts a notice in the meeting chat. Teams can ask questions by voice or chat, and Claude responds in writing by default while also creating documents, whiteboards, and other artifacts during the call. Each meeting has a dedicated page in claude.ai containing notes, a transcript, and generated content. Invited members of the organization can view the content and edit the artifacts. The system saves transcripts, summaries, and screen descriptions, but not audio.
Inviting Claude is enabled by default on Team plans and can be disabled by an admin. On Enterprise plans, an admin must enable it for all or part of the organization. By default, Claude stays out of meetings with invited external guests and leaves if an external participant joins midway, although admins can allow exceptions. Claude requests permission before using a person's connected tools, and the initial approval remains valid for subsequent use. For now, spoken requests still require confirmation of the speaker each time. During the beta, notes, transcripts, and meeting answers that do not use connected tools are free; answers using tools count toward the plan's usage. Documents, whiteboards, and other artifacts are free through February 10, 2027. Some organizations using zero data retention (ZDR), HIPAA-ready configuration, or customer-managed encryption keys (CMEK) cannot currently access the beta.
Grok Bot for Mac integrates with seven app categories, including messaging and calendars
要闻
Grok Bot announced that its Mac app can now use 7 apps and features, including Messages and Mail, adding context and tools for the bot. The team cited organizing important family events and helping small businesses manage appointments as example uses.
Grok Bot announced that its Mac app can now use 7 apps and features: Messages, Mail, Notes, Reminders, Calendar, Location, and Contacts. Team member lauren said that adding context and tools enables the bot to help users with more work and everyday tasks. She outlined 2 use cases: checking family messages and adding important events to the calendar, and reviewing messages from a small business’s prospective customers to help manage appointments. These scenarios are examples of uses provided by the team.
Thinking Machines has released Inkling 2 Small weights on Hugging Face under the Apache 2.0 license. The multimodal MoE model has 594B total parameters and 17B active parameters, with the NVFP4 version requiring at least 500GB of GPU memory.
Thinking Machines has made the Inkling 2 Small model weights available on Hugging Face under the Apache 2.0 license; the source does not specify a release date. The model uses a multimodal MoE architecture with 594B total parameters and 17B active parameters, supporting image and text inputs and a context window of up to 1M tokens. According to the company, the weights come in NVFP4 and MXFP8 precision, with the NVFP4 version requiring at least 500GB of GPU memory. For deployment, vLLM added support on the day the weights were released, integrating a vision encoder, MoE expert parallelism, and DSpark speculative decoding, alongside LoRA support. The model can also run through frameworks including SGLang and Unsloth, or be accessed through APIs on platforms including NVIDIA and Together.
IBM introduces LLMON markup language to reduce prompt injection risks
模型发布
IBM released LLMON and two experimental preview models based on Granite 4.1, using explicit structure to separate instructions from data. According to IBM, they avoided 47% more attacks on OPI than the best frontier model tested.
IBM released LLMON (LLM Object Notation) and two experimental preview models, granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview; the source does not specify a release date. Both models derive from Granite 4.1 base models and can consume and generate LLMON text. An accompanying GitHub repo provides a toolkit and examples of workflows and templates. LLMON marks structural boundaries with reserved special tokens and uses unique IDs to bind instructions to the corresponding data, aiming to reduce the risk of indirect prompt injection through instructions hidden in data. It does not replace chat templates, and tools also support lossless conversion between JSON and LLMON. According to IBM, on the OPI benchmark measuring static indirect prompt injection, the preview models avoided 47% more attacks than the best frontier model tested and 95% more than the best open-source model. IBM says their general capabilities are on par with Granite 4.1 Instruct models, while defenses against adaptive attacks spanning multiple conversation turns remain under development.
OpenAI launches two AI-powered cybersecurity defense tools
开发生态
OpenAI announced two cybersecurity tools: Daybreak Red Team offers early access to selected customers and has opened a waitlist, while Daybreak Defense Factory scans code and handles vulnerabilities as a research preview.
OpenAI announced Daybreak Red Team and Daybreak Defense Factory during its Cyber at the Frontier keynote, with the former available in early access to selected customers and the latter currently in research preview. According to the company, the tools cover application testing and code security workflows, respectively. Red Team consists of collaborating security agents that test Web applications, verify whether vulnerabilities can be exploited, and check whether patches work; its waitlist is open. Defense Factory is built on Codex Security and scans code undergoing changes, reviews pull requests, and investigates alerts. When it finds a vulnerability, it reports it to the responsible engineer and proposes a patch in Codex. It scans connected GitHub repositories through the Daybreak plugin, supporting one-time scans, scheduled scans, and continuous monitoring of new commits.
Claude API introduces a beta feature for adjusting effort mid-conversation
开发生态
Claude API documentation lists mid-conversation effort changes for 5 models, with per-message configuration preserving the prompt cache. Availability is subject to platform and thinking-mode restrictions, and the source does not specify a launch date.
Claude API documentation now lists the ability to change effort during a conversation, without specifying a launch date. According to the official documentation, per-message output_config settings preserve the prompt cache when effort changes. The capability covers Claude Fable 5.2, Claude Mythos 5.2, Claude Opus 5.5, Claude Sonnet 5.5, and Claude Haiku 5.5. The descriptions for Fable, Mythos, Sonnet, and Haiku explicitly limit availability to supported platforms. Effort is set through output_config.effort in the request and controls token expenditure across the response, including thinking and tool calls.
The capability is constrained by the thinking mode. According to the official documentation, when Claude Sonnet 5.5 uses between_tools, a per-message configuration that changes effort to a different level returns a 400 error; varying effort by turn requires adaptive thinking. Claude Haiku 5.5 also cannot change effort mid-conversation when thinking is disabled. Defaults differ by model as well: Claude Opus 5.5 and Claude Haiku 5.5 use medium, while Claude Fable 5.2, Claude Mythos 5.2, and Claude Sonnet 5.5 use high.
Managed Agents launches a private beta for enterprise self-hosting
开发生态
Anthropic has opened a private beta of self-hosted Claude Managed Agents, allowing enterprises to run the entire service in their own cloud or on-premise rather than hosting only the sandbox. Developers retain the hosted version's core APIs, while enterprises can choose where model requests are routed.
Anthropic has made self-hosted Claude Managed Agents available in private beta, enabling enterprises to deploy the complete service in their own cloud or on-premise environment. Previously, Anthropic operated the infrastructure on Claude Platform and later added a self-hosted sandbox option. The new deployment scope includes the harness, sandboxes, credentials, and session history, with enterprises operating the Managed Agents server, database, and workers where agents execute shell and file tools. Developers continue to use the same core APIs as the hosted service, and model requests can be sent to Anthropic, Amazon Bedrock, Google's Agent Platform, Microsoft Foundry, or an enterprise's own gateway.
According to the company, the security architecture follows four principles: separating the harness from the sandbox, keeping credentials outside the sandbox, restricting network access, and retaining complete session records. The agent loop, session history, and vault secrets all remain outside the sandbox. An external proxy adds credentials to MCP tools and supported API calls, and the self-hosted version also supports an enterprise's own secrets manager. Enterprises can configure allowed hosts and enforce restrictions for self-hosted sandboxes through their own firewall. Each session's tool calls and results can be exported through OpenTelemetry. Session history is stored in the enterprise's database, and existing zero data retention (ZDR) arrangements with Anthropic also apply to requests sent to Claude. Enterprises can contact their Anthropic account team to request beta access.
Google releases A2UI 1.0, an open protocol for generative UI
开发生态
Google has released A2UI 1.0, the first stable version of its open protocol for generative UI, enabling Agents to respond with interactive interfaces. It provides open-source renderer SDKs for web, Android, and iOS, and multiple companies are already using it.
Google has released A2UI 1.0, its open protocol for generative UI; the source does not specify the release date. This is the protocol’s first stable version, accompanied by open-source renderer SDKs for web, Android, and iOS and agent SDKs for Python, TypeScript, and Dart. The Agent describes the interface content, while the application draws it using its own native components, allowing components from different applications to be combined in one view. According to Google, this mechanism lets the same Agent output render natively across platforms while maintaining low latency and security. A2UI can also work with MCP Apps, and Alation already uses both to display data feeds containing metric cards, charts, and custom components inside Claude.
Supported protocol versions differ across products. ByteDance released an A2UI renderer for its cross-platform framework Lynx in June 2026. It currently supports A2UI 0.9 and is being adapted to 1.0 in collaboration with the A2UI team. Gemini currently supports A2UI 0.9.1; organizations can register Agents that return A2UI responses, which Gemini renders using its component gallery. Google says Jetpack Compose support for 1.0 is still forthcoming and Gemini Enterprise is being upgraded to A2UI 1.0. As an application example, Atlassian is bringing A2UI to Jira so that Agents can request additional information or decisions through interactive cards within work items, allowing team members to respond directly in Jira.
Google has announced that pre-orders are open for Gemini-powered smart glasses from Gentle Monster and Warby Parker, starting at $449 and $395, respectively. Both collections support hands-free interaction and work with Pixel and Samsung phones as well as iPhones.
Google has opened pre-orders for smart glasses developed with Gentle Monster, Warby Parker, and Samsung, with ordering and try-on information available through the brands’ websites or Google Store. The products come in two collections: Gentle Monster offers six designs starting at $449, with sales planned in the United States and South Korea and prescription lenses available for most frames. Warby Parker offers three styles—Alden, Reid, and Wren—starting at $395, with optical and sunglasses options, along with a choice of frame widths and lenses. Warby Parker also offers guided experiences by appointment at more than 360 stores.
According to the official description, both collections integrate Gemini, work with Pixel and Samsung phones as well as iPhones, and can be operated by voice or by touching the frames. Features include Google Maps audio navigation, real-time speech translation across more than 70 languages with automatic language detection, and conversational audio descriptions through Guided Vision. Users can also invoke Aira to send the camera feed to a professionally trained human visual interpreter. Official hardware specifications include a weight of around 41 grams, up to nine hours of battery life, a portable case capable of recharging the glasses up to seven times, and a 12-megapixel ultra-wide camera. The glasses also support photos, video recording, music playback, message replies, and hands-free calls.
Google introduces five new artist-designed Flow tools
产品应用
Google announced five creator-built tools for its AI creative studio Flow, covering characters across different ages, image style conversion, glass lighting effects, and phone video processing. According to Google, users can also build their own creative tools in the studio.
Google announced the release of five creative tools in Flow; the source does not specify a release date. For character design, James Gerde’s Character Poster turns story concepts into visual posters to help maintain a consistent style across social content, while Anurag Tiwari’s Character Arc depicts the same character at different ages and life stages. According to Google, Tiwari is using the latter in the documentary “Children of the Sun” to depict the protagonist’s progression from childhood to adulthood. For image processing, Tiwari’s Pixel Machine converts images into bitmap, ASCII, CRT, pixel art, and glitch styles, while Ezra Paulekas’s Lucid Glass generates glass lighting effects such as prisms, halos, and refractions. For video, Ryan McCoy created Pocket Studio, which Google says can generate multiple camera angles from a single phone take and improve visual quality while preserving the actor’s original performance.
OpenRouter website launches a beta version in seven languages
产品应用
OpenRouter launched its website localization beta on November 12, 2026, adding seven languages, including Japanese and Simplified Chinese, for a total of eight with English. According to the company, a daily automated pipeline updates translations, with a person merging them for release.
OpenRouter released its website localization beta on November 12, 2026, adding seven languages: Simplified Chinese, Spanish, Brazilian Portuguese, German, French, Japanese, and Korean. Users can switch languages through the account menu or visit URLs such as openrouter.ai/ja directly. According to the company, the site contains 16,918 English strings, with changes made every day. Translations are stored alongside the code as JSON catalogs and loaded locally at runtime, keeping translation services out of the request path. Content that has not yet been translated appears in English.
According to the company, the daily pipeline identifies new or changed strings and sends them to GPT-6.1 Sol in batches of up to 40, using Structured Outputs to constrain the returned JSON. After code checks formatting and terminology, a separate call to the same model compares the translation’s meaning with the English, triggering retries for serious errors. If the verifier model is unavailable, the pipeline writes no translations and the content remains in English. Each daily update becomes a PR, with an AI reviewer examining the diff before a person merges it. In the company’s test of 254 cases, GPT-6.1 Sol achieved a first-pass success rate of 99.2% at a cost of $1.94; the corresponding results for GPT-6 Luna were 91.7% and $0.57.
Russian prankster duo dupes OpenAI by posing as Ukraine’s foreign intelligence chief
行业动态
Russian pranksters impersonated Ukrainian officials in a 17-minute video call with three senior OpenAI military and national security staff, discussing military AI support for Ukraine and offensive uses. OpenAI said it had investigated and taken preventive measures.
The Intercept reported on November 11, 2026, that Russian duo Vovan and Lexus had impersonated Ukrainian officials to hold a video meeting with three senior OpenAI staff; the specific date of the call was not disclosed. The participants were Sasha Baker, who oversees national security policy; Caroline Zier, who leads strategic initiatives; and Ben Crampton, who handles European national security matters. The duo said they secured the meeting by posing as Rustem Umerov, head of Ukraine’s Foreign Intelligence Service, and uploaded a recording of the 17-minute call to Rumble.
According to the recording, the discussion included the possibility of providing offensive military AI capabilities to Ukraine. Crampton said some senior figures at the company were reluctant to become deeply involved in offensive military uses, but that offense and defense could not necessarily be distinguished in some circumstances. Baker said the company had contacted the Pentagon about assisting Ukraine. OpenAI staff also discussed work with the Ukrainian government involving Daybreak Blue AI model, a system for fixing computer vulnerabilities, and an early-stage partnership with Ukraine’s Ministry of Defense. OpenAI spokesperson Nate Evans said the company had confirmed through an internal investigation and cooperation with relevant authorities that the call was a prank, and had taken steps to prevent similar incidents.
腾讯 denies rumors of 50% layoffs, calling them a misinterpretation
行业动态
Tencent public relations director Zhang Jun denied rumors that management planned to cut 50% of staff, saying the figure came from an analogy about roles at an all-hands meeting. According to a person close to Tencent, the meeting outlined a split in which 50% would handle existing core businesses and the other 50% would explore AI Native.
Tencent public relations director Zhang Jun, responding to rumors that management planned to cut 50% of staff, said the claim arose from a misunderstanding and was untrue. He explained that the 50% mentioned at an internal all-hands meeting referred to a division of roles likened to an “army” and a “navy,” intended to describe the reallocation of resources for the AI era rather than a layoff percentage. According to a person close to Tencent, management proposed accelerating organizational changes at the meeting, with 50% of staff working on existing core businesses through efficiency improvements and the other 50% exploring AI Native and new areas for growth.
张一鸣 says 字节’s AI will catch up with the frontier in 2029 or 2030 at the earliest
行业动态
ByteDance founder Zhang Yiming told its AI team that the company can compete with OpenAI and Anthropic, but will not catch up with frontier technology until 2029 or 2030 at the earliest, according to The Paper. He also identified a gap in the scale of Seed’s pretraining datasets.
ByteDance founder Zhang Yiming put the earliest date for the company to catch up with frontier AI technology at 2029 or 2030 during an AI team all-hands meeting whose date was not disclosed, according to The Paper, citing multiple people familiar with the matter. The report said he believes ByteDance can compete with OpenAI and Anthropic and should benchmark its large models against the international top tier, but that the Seed series’ pretraining datasets remain smaller than those of leading overseas AI companies, making this the most pressing issue. Separately, overseas media reported that the Seed team is pretraining a model with up to 10 trillion parameters, comparable in scale to Anthropic’s highest-end model. The Paper previously cited ByteDance insiders as saying that Zhang had called at an internal meeting for sustained, long-term investment and delayed gratification in model development, rather than using other models’ outputs to secure short-term leaderboard positions.
百度 launches large-model marketplace “词元所,” touting neutrality and trustworthiness
行业动态
Baidu announced Token Exchange, a Token trading platform connecting model suppliers and users, with a unified API for accessing and switching models. According to Baidu, average daily Token usage exceeded 700 billion within the platform’s first month.
Baidu announced the launch of “词元所” (Token Exchange), a Token trading platform connecting large-model suppliers and users, at Baidu World 2026. The platform offers comparisons of model pricing, performance, latency and costs, alongside a unified API for model access and switching. Baidu reported that, over the first month after launch, the platform attracted more than 100 supply resources, with average daily calls exceeding 50 million and average daily Token usage exceeding 700 billion. On resource management, the company said it would conduct admission tests and post-launch spot checks, and create publicly accessible credit records for model providers.
