AI News

AI news brief: Wednesday, November 11, 2026

26 stories, most important first.

ReportedSafety & Security · · Updated

Anthropic's S-1 says Autonomy 2 threshold met under accelerated R&D condition

Anthropic's updated S-1 reportedly states that its Autonomy 2 capability threshold has been met, and says the finding does not rest on current models.

Why it matters: A stated capability threshold for autonomous AI research in a public filing is a governance signal. Customers and regulators may cite it when assessing frontier model risk.

2 sources
OfficialModels · · Updated

Google releases Gemini Decision model and Decisions API for binary and categorical calls

Google's Gemini Decision model and Decisions API return binary, categorical and rubric decisions for text, images, audio, video and PDFs, and Google says they use zero output tokens.

Why it matters: If the zero-output-token claim holds for your workloads, teams that now generate text just to get a label or score could cut cost and latency. Buyers should check the published pricing and limits before relying on it.

4 sources
OfficialOpen Source · · Updated

Thinking Machines unveils Inkling 2 Small model, with weights due next week

Matthew Berman says Inkling 2 Small scores 31 on the Artificial Analysis Intelligence Index and can be downloaded, but the company says the weights are not yet posted.

Why it matters: A compact open-weights model with a reported strong score gives teams a cheaper option for running smaller agent tasks.

2 sources
ReportedBusiness & Funding · · Updated

Mercor raises $750M Series D at $20B valuation on $3B run-rate revenue

Mercor raised a $750 million Series D at a $20 billion valuation, according to a post by Brendan Foody, after reaching $3 billion in annualized revenue.

Why it matters: Mercor's growth shows that demand for expert-labeled data and model evaluation is still rising fast. Buyers of these services should expect pricing and capacity to shift quickly.

2 sources
OfficialChips & Infrastructure ·

Anthropic says Claude runs on four chip families, with AMD port done in three days

Anthropic's IPO roadshow video says Claude now runs on four chip families, and that an AI-driven AMD port took three days.

Why it matters: If the claims hold, Anthropic can move Claude across several accelerator vendors, which could widen the hardware options behind the models buyers rely on. The figures come from company statements in an IPO roadshow video and are not independently verified.

1 source
ReportedSafety & Security ·

Railway says hackers stole credentials for 100,000 workspaces

Hackers exploited Railway's build infrastructure to steal credentials for 100,000 workspaces, then ran up $740,000 in fraudulent AI charges.

Why it matters: Teams that run AI workloads on hosted platforms face direct financial exposure if stolen credentials are used to run paid model calls.

1 source
ReportedPolicy & Regulation ·

Report: Shanghai Century Litong facilitated at least $7.5B in Nvidia chip sales

The Information reports that Shanghai Century Litong, a company tied to Alice Huang, facilitated at least $7.5 billion in Nvidia chip sales to China.

Why it matters: Enforcement against chip diversion can tighten export compliance across the supply chain, including for resellers and cloud providers selling into China. Buyers of AI hardware should expect closer scrutiny of end users.

1 source
ReportedOpen Source · · Updated

Recursive Intelligence releases open-weights Ripple 65B preview model

Commentator Matthew Berman says the 65B dense Ripple model supports a 1M-token context and scores 80.5% on Terminal Bench 2.1 among open-weights models.

Why it matters: A small open-weights model with long context and low quoted prices gives builders a self-hostable option for agent workloads.

2 sources
OfficialProducts ·

Codex CLI 0.167.2 sends tool updates incrementally by default on Vertex AI

Codex CLI 0.167.2 sends tool updates incrementally by default on Google Cloud Vertex AI, with a config setting to opt out.

Why it matters: Teams running Codex CLI on Vertex AI will see a different tool-update pattern after upgrading, so anyone with tooling that depends on the old behavior should check it.

1 source
OfficialAI in Practice ·

AWS releases aws-emulator for local tests, written in Rust by a swarm of agents

AWS has released an official emulator that lets developers run tests against a local copy of its services without a paid AWS API key.

Why it matters: Teams that test AI and cloud workloads against AWS services could run those tests locally instead of paying for live API access. The agent-built codebase is also a concrete example of agentic engineering at scale.

1 source
ReportedAI in Practice ·

Duolingo says AI agents merged 700+ test PRs, lifting iOS coverage from 11% to 24%

Duolingo engineer Kush Agrawal says an agent pipeline has merged more than 700 unit-test PRs into the iOS app, and a human approved each one.

Why it matters: Teams adopting coding agents get a concrete case in which verification, not generation, became the bottleneck. Automated tests and review gates were needed before agent output could be merged safely.

1 source
OfficialResearch ·

Stanford funds second round of AI and neuroscience grants for five research projects

Stanford's Wu Tsai Neurosciences Institute and Institute for Human-Centered AI have funded five projects that link brain research and AI.

Why it matters: This is early-stage academic research with no product changes. Its most relevant idea for buyers is that the infant-learning project aims to make AI systems learn from less data more efficiently.

1 source
ReportedBusiness & Funding ·

Bay raises $37M Series A for real-time checks on AI agent actions

Bay's real-time enforcement layer checks agent actions before they execute across endpoints and cloud, and the company says it serves Fortune 100 customers.

Why it matters: Enterprises moving agents into production may need a control layer that checks actions before they run, not only after.

1 source
ReportedProducts ·

OpenAI releases GPT Image 2.6 with GIF generation and faster image output

Matthew Berman reports that OpenAI's GPT Image 2.6 can generate GIFs and runs about 50% faster, in two variants called Sunburst and Flare.

Why it matters: Companies generating images in products get a faster option and a choice between quality and speed, which affects cost and latency planning.

1 source
ReportedProducts ·

Google releases Gemini Omni 1.2 Flash video generation model

Matthew Berman says Google's Gemini Omni 1.2 Flash is faster and about 30% cheaper than its predecessor, with longer clips and in-chat editing.

Why it matters: Teams building video features may get lower generation costs and longer outputs from a faster model.

1 source
ReportedOpen Source ·

LongCat 2.5 open model reaches 1.6 trillion parameters with 48B active

Matthew Berman says LongCat 2.5 is a 1.6-trillion-parameter multimodal model with 48 billion active parameters that can be downloaded.

Why it matters: Very large open models widen the choice for organisations with heavy inference infrastructure, though most teams will need hosted access.

1 source
ReportedModels ·

MiniMax releases open-weights H3 Image model, 68 GB or 18 GB quantized

Matthew Berman says MiniMax's H3 Image is a fully open-weights image model, 68 GB at full size and about 18 GB in a quantized version.

Why it matters: A downloadable image model that runs on consumer hardware lets teams keep image generation in-house and avoid per-image API fees.

1 source
ReportedModels ·

ByteDance releases Seedream 5.5 image model with multi-turn editing

Matthew Berman calls ByteDance's Seedream 5.5 a top image generator and says it supports multi-turn editing that keeps images consistent.

Why it matters: Multi-turn consistent editing matters for product and marketing teams that need several revisions of the same image.

1 source
ReportedProducts ·

Anthropic gives Claude paid users one free usage reset after limit bugs

Anthropic is giving Pro, Max, Ultra and Team users one free full usage reset, usable until December 10, after fixing bugs that consumed extra limits.

Why it matters: Teams that depend on Claude plans for daily work get a short window to recover capacity. The bugs also show that usage can run out faster than expected.

1 source
OfficialOpen Source ·

Model Context Protocol moves to two releases a year, next due March 30, 2027

MCP specification releases move to a regular six-month schedule, with the next version targeted for March 30, 2027 and proposal deadlines published in advance.

Why it matters: Teams building on MCP can plan proposals, SDK upgrades, and testing against fixed dates instead of informal meeting notes. The schedule also means features that miss a deadline will wait six months.

1 source