AI News

AI news brief: Thursday, November 12, 2026

36 stories, most important first.

OfficialModels · · Updated

Google DeepMind introduces Gemini 4 Hydrogen

Google DeepMind announced Gemini 4 Hydrogen in a company post that reports benchmark results and pricing, though those figures are Google's own.

Why it matters: Teams that plan model choices around Google's roadmap should wait for the capability and pricing details before changing their plans.

4 sources
ReportedChips & Infrastructure · · Updated

Former OpenAI researcher says banned Blackwell chips trained top Chinese AI models

A former OpenAI researcher says freely released models from China's top labs carry a hidden fingerprint of Nvidia Blackwell chips that were banned from China.

Why it matters: If the claim holds, it would suggest that restricted hardware reached Chinese labs and shaped open models, which could draw tighter export enforcement and closer scrutiny of where AI models come from.

3 sources
ReportedBusiness & Funding · · Updated

WSJ reports Nvidia board preparing to remove Jensen Huang as CEO

The Wall Street Journal reports Nvidia's board is preparing to remove Jensen Huang, following a New York Times report of a second affair.

Why it matters: A leadership change at Nvidia, the main supplier of AI accelerators, could affect buyers' chip roadmaps and supply planning, though the reports are unconfirmed.

3 sources
ReportedBusiness & Funding · · Updated

OpenAI quarterly revenue crosses $10 billion, up 50% from Q2

OpenAI's quarterly revenue crossed $10 billion for the first time, up 50% from $6.7 billion in Q2, as enterprises adopt its models and agent tools.

Why it matters: Strong enterprise demand suggests OpenAI's vendor position is growing, which matters for buyers weighing long-term supplier risk.

2 sources
ReportedSafety & Security · · Updated

Fireworks AI breach exposed Notion AI chat transcripts

A breach at Fireworks AI exposed Notion AI chat transcripts, reaching under 1% of Notion workspaces despite zero data retention claims.

Why it matters: Vendors' data retention claims may not cover stored batch inference data, so buyers should verify where prompts and outputs are kept.

2 sources
OfficialProducts ·

Google brings Gemini 4 Hydrogen into Google Search as a major upgrade

Google says Gemini 4 Hydrogen is now available in Google Search, which it calls a major upgrade to the product.

Why it matters: Google Search is a major traffic source, so changes to the model behind it can alter how businesses are found and how their content is surfaced.

1 source
ReportedModels · · Updated

Claude Fable 5.2 ranks #1 on Agent Arena; API price falls to $8/$40 per million

Arena ranks Claude Fable 5.2 (Max) first with a +30.4% net improvement across 4.7K agentic sessions, ahead of Opus 5.5 at +19.5%.

Why it matters: The lower token price comes with a higher median cost per task than Opus 5.5. Buyers should compare cost per completed task rather than price per token.

1 source
ReportedModels ·

Runway releases Praxis-1 world action model with Apache 2.0 weights

Runway's Praxis-1 is a 27B MoE world action model, and Runway says it finished robot tasks 15.6x faster than GPT-6 Astra.

Why it matters: Open Apache 2.0 weights give robotics teams a self-hostable alternative to closed world models, although the speed and cost gains are company-reported.

1 source
ReportedModels ·

Anthropic says Fable 5.2 and Mythos 5.2 were capacity-capped in AI R&D training

A document excerpt shared on Reddit says Fable 5.2 and Mythos 5.2 were deliberately capacity-capped during training in AI R&D domains.

Why it matters: If training caps reduce skill in kernel or systems work, teams relying on Claude for performance engineering may see uneven results that narrow benchmarks would not reveal.

1 source
OfficialModels · · Updated

Google makes Gemini 4 Hydrogen generally available at $1.50 and $7.50 per million tokens

Google has made Gemini 4 Hydrogen generally available with a new xhigh thinking level and named it the default model for the Antigravity agent.

Why it matters: Teams on Gemini 3.8 Flash now have a newer GA model, a documented migration path, and a stated price, which matters for cost planning on coding agents. The xhigh level gives builders one more reasoning-cost setting to test.

1 source
OfficialProducts · · Updated

Claude Code now assumes a 1M-token context for unrecognised model IDs

Claude Code 2.1.323 raises the context window it assumes for unknown model IDs from 200K to 1M tokens.

Why it matters: Teams routing non-Anthropic models through Claude Code may hit failed turns if their models have smaller windows than Claude Code now assumes.

1 source
ReportedChips & Infrastructure ·

Torus raises $500M to build modular inference infrastructure

Torus raised $500 million in growth funding to deploy modular inference capacity of 1 to 20 MW in 12 to 16 weeks, targeting 47,000 GPUs by end of 2027.

Why it matters: Faster inference capacity with existing grid connections could ease power and deployment constraints for companies running AI in production.

1 source
ReportedAI in Practice ·

Claude Code team lets low-risk PRs merge on automated checks alone

The Claude Code team sets human approval by each change's risk and scores its AI reviewer daily against PRs that caused past incidents.

Why it matters: Risk-tiered approval shows how teams can use AI code review to speed up low-risk merges while keeping human checks on riskier changes.

1 source
ReportedSafety & Security ·

OpenAI commits $2 million to research on AI and children's safety

OpenAI says it is committing $2 million to research on AI and children's safety, aimed at shared tools that measure risks and opportunities for young people.

Why it matters: Companies that build or deploy AI products used by minors could eventually use common measurement tools for child-safety risks. The announcement gives no timeline and no details on what the tools will do.

1 source
OfficialSafety & Security · · Updated

Claude Code 2.1.323 closes several Bash permission and sandbox gaps

The Claude Code 2.1.323 changelog lists fixes for Bash rules, Write and Edit paths, and sandbox behaviour, and a change to /security-review.

Why it matters: Companies that rely on Claude Code permission rules to limit what agents can do should know that several of those rules had gaps, and that some commands will now ask for approval.

1 source
OfficialProducts · · Updated

Claude Code removes subagentStatusLine and deprecates a plugin agent setting

Claude Code 2.1.323 removes the subagentStatusLine setting and deprecates the plugin agent setting, which now triggers a startup warning.

Why it matters: Plugin authors who rely on the agent setting will need to migrate before it stops working. Teams with custom subagent displays will lose that option.

1 source
OfficialProducts ·

Claude Code drops the v1 MCP client and runs v2 in compatibility mode

With the v1 MCP client removed, the MCP_SDK_GENERATION=v1 setting now runs the v2 client with an earlier handshake and relaxed OAuth checks.

Why it matters: Teams whose MCP servers depend on the older client behaviour may see changed connection or sign-in behaviour.

1 source
OfficialProducts ·

Claude Code raises cross-session message limit from about 20 to 100,000

Claude Code 2.1.323 raises how many messages a session can send to other sessions on the same machine without a prompt, from about 20 to 100,000.

Why it matters: Agents coordinating across sessions can now run far longer without human checkpoints, so errors they repeat can spread further before anyone steps in.

1 source
OfficialProducts ·

Claude Code plugins can declare a version range and skip out-of-range installs

Claude Code 2.1.323 lets plugins state a supported Claude Code version range in plugin.json, and plugins outside that range are not installed, updated or loaded.

Why it matters: Plugin authors can avoid breaking users on incompatible Claude Code releases, but teams need to confirm their plugins declare a valid range.

1 source