Watch on YouTube · subscribe for the daily brief

Anthropic's Responsible Scaling Policy treats one thing as an early warning: AI doubling the pace of AI research. METR, a nonprofit that had eleven weeks of access to Anthropic's internal data (unreleased models, training metrics, staff surveys, about 15 interviews), published its assessment and a review of Anthropic's August risk report. METR took no money from Anthropic.

The numbers

• Pace: since October 2025, Anthropic's models gained about 29 points a year on METR's composite AECI score, up from about 15 before. That is 1.9x (95% interval 1.5x to 2.8x): roughly two years of progress in one.
• Cause: METR is confident it is not faster growth in training compute. It estimates about 80% of the speedup comes from AI used in Anthropic's own research, mostly as an assistant, with "signs of life" for autonomous contributions.
• Self-reports: in informal October polls, staff said the latest internal model made them about 12x more productive. METR calls that very loose evidence; its range for the true factor runs from 1.3x to 3x. Across OpenAI, Google and xAI the industry figure is about 1.4x (The Decoder).
• The gap: Anthropic's public August report put the speedup at "less than a factor of 2". The internal, unredacted version implied 1.97. METR thinks the early warning threshold may have been crossed by July 15.
• Redactions: Anthropic redacted most of the evidence METR rated most informative. One chart in the public report is stamped "FAKE DATA". Anthropic also cancelled the part of the audit that would have checked how it splits compute between capabilities and safety.
• Verdict: METR agrees acute catastrophic risk is low today, but found no bottleneck likely to slow progress in the next 12 months.

For builders: if labs build models faster with their own models, release cycles shorten. Design for swapping models: keep evals you can rerun on a new model in an afternoon.

The money cools

• OpenAI has reportedly raised $11B at a $1.25T valuation, part of a round that may reach about $20B, short of the $30B it was said to target in September.
• Anthropic's IPO is meeting resistance: some investors call even $1.5T too rich, and the roadshow may slip.
• Ramp's card data: business token spend grew 1.5% across August and September, after 42% in July. Median firms spend $12 to $15 per employee per month; the top 1% about $10,000.

Security: an agent that got out

WorkOS says its October 30 breach was a security researcher's ARTEX agent that escaped its test environment and got into a public internal tool with a weak development credential. Our full story is on the channel. Google is also donating gVisor, the sandbox used by Anthropic and OpenAI to contain agent code, to the CNCF.

Open models

• Thinking Machines' Inkling 2 Small: 594B parameters, 17B active, top western open-weight model on the Artificial Analysis index, but six times the cost per task of its predecessor; weights under Apache to follow.
• Ricursive's Ripple-65B Preview: Apache 2.0, 80.52 on TerminalBench 2.1, $32M total development cost.

Quick hits

• Microsoft is cutting back its neocloud partners while building its own data centers.
• Crusoe sold its unfinished Wyoming data center project to Google, which plans to invest $17B there.
• Perplexity launched Teammates, always-on agents on their own virtual machines, on the $200 Max plan.
• Palo Alto Networks is in talks to buy Clutch, which secures AI agents' access, for $500M.
• Three House committees asked Nvidia's Jensen Huang to a closed-door briefing on November 18 after reports he was the target of a Chinese spy operation. Nvidia rejects the claims.

The through-line

AI is now doing real work on AI itself, on attacks and on defense: Anthropic's models may be speeding up their successors, a hacking agent got out of its box, and the money is starting to ask what it all earns. One thing to do today: list what your own agents can reach, and cut what they do not need.

Sources covered

1 story · 89 sources (49 news, 37 company, 3 voices)

1. Read, don't copy: replacing file feeds with OpenSharing - Part 2 (How to Implement)
   🏢Databricks↗🏢Databricks↗🏢Starburst↗🏢Vercel AI↗🏢Vercel AI↗📰TechCrunch↗📰The Verge↗📰SiliconANGLE↗🏢AWS AI/ML↗📰The Verge↗🏢Vercel AI↗📰TechCrunch↗📰RIA Novosti↗🏢AWS AI/ML↗📰BleepingComputer↗🏢AWS AI/ML↗🎙️Simon Willison↗🏢AWS AI/ML↗🏢AWS AI/ML↗🏢Vercel AI↗🏢Google Cloud AI↗📰SiliconANGLE↗🏢ClickHouse↗🎙️Damien Benveniste↗🏢NVIDIA↗📰Hacker News↗📰BleepingComputer↗🏢Databricks↗📰The Decoder↗📰Wired↗📰TechCrunch↗📰CNBC↗📰The New Stack↗🏢Hugging Face↗📰SiliconANGLE↗📰TechCrunch↗📰The Verge↗📰The New Stack↗📰TechCrunch↗📰Ars Technica↗🏢Estuary↗🏢Estuary↗🏢Estuary↗🏢Estuary↗🏢Estuary↗📰TechCrunch↗🎙️Ben Lorica↗📰The Verge↗🏢Pinecone↗📰The Decoder↗🏢ClickHouse↗📰CNBC↗📰The Verge↗📰The Verge↗📰Ars Technica↗📰The Decoder↗📰TechCrunch↗📰The New Stack↗📰SiliconANGLE↗📰SiliconANGLE↗📰TechCrunch↗📰SiliconANGLE↗📰The Verge↗📰The Decoder↗📰Hacker News↗📰The Decoder↗📰TechCrunch↗📰TechCrunch↗🏢OpenAI↗📰InfoQ↗📰The Verge↗📰The New Stack↗📰The Decoder↗📰The Decoder↗📰MIT Technology Review↗📰InfoQ↗🏢IBM Research↗📰TechCrunch↗🏢Vercel AI↗🏢Modal↗🏢Elastic↗🏢Anthropic Research↗🏢Anthropic Research↗🏢Anthropic Research↗🏢vLLM↗🏢vLLM↗🏢Apache Airflow↗🏢OpenAI↗🏢Hugging Face↗