Technology · news

METR finds faster AI research at Anthropic. The evidence is only partly public.

An independent assessment estimates roughly twice the historical pace on well-defined tasks, while redactions limit outside scrutiny.

By UpThatNov 6, 2026, 21:08 WIB3 min read4 sources

Ink illustration of a hand marking an open research notebook beside a page covered with broad black redaction bars. — Conceptual illustration of AI research evidence and restricted disclosure. The notebook marks and symbolic graph are invented; this does not reproduce METR’s report or measurements. Generated with GPT Image for UpThat AI-generated editorial illustration.

Anthropic’s progress on well-defined AI tasks is likely running at roughly twice its March 2024–October 2025 pace, independent evaluator METR reported on November 5. Its assessment attributes much of the acceleration to AI assistance, with substantial uncertainty. Anthropic redacted most of the evidence METR considered most informative.

The measurable anchor is AECI, Anthropic’s adaptation of the Epoch Capabilities Index. METR estimated that benchmark-score growth since October 2025 accelerated 1.9 times, with a 95% confidence interval of 1.5–2.8. That compares rates of improvement; it does not mean a model is twice as intelligent.

METR examined internal information from August 12 to October 29. It did not receive access to sample unreleased models. Researchers described assistance rather than fully autonomous contributions to core training work. Benchmark gains may not transfer to messier tasks. This assessment was separate from October’s employee-like-access arrangement.

An early warning is a different question from immediate danger

A companion METR review examines Anthropic’s August risk report, covering models available by July 15. METR broadly agrees that their direct catastrophic risk from automated research was low. It challenges how confidently the public report ruled out crossing an early warning threshold.

That threshold concerns a doubling of AI progress substantially attributable to research automation, rather than simply faster individual researchers. METR says crossing it by July 15 was plausible, although most relevant researchers at both organizations leaned toward it not having been crossed. Those judgments concern the earlier period, not October’s capabilities.

The distinction matters: a warning threshold is intended to prompt scrutiny before a feared outcome materializes. The review does not establish that models could replace Anthropic’s entire research workforce, and agreement about low immediate risk does not settle whether an earlier warning deserved greater emphasis.

Access for evaluators does not guarantee visibility for readers

In its October 15 announcement, Anthropic described onboarding six METR staff with hardware, office space and internal tools roughly comparable to its risk teams’ access. Exceptions included customer data, legally privileged material and model architecture information. That announcement describes a separate engagement.

Anthropic says it cannot veto that team’s publication but can require redactions under predefined confidentiality categories, whose application it ultimately decides. METR can disclose aspects of access and information received. The arrangement therefore separates an evaluator’s ability to investigate from the public’s ability to inspect the underlying evidence.

Responding to the new findings, Anthropic co-founder Sam McCandlish said many conclusions matched his own, according to the Associated Press. He acknowledged that AI was speeding research while the size of the effect remained hard to determine. This is the company’s response, not an independent measurement.

For readers assessing the next claim of AI-driven acceleration, three questions now carry more weight: what improved, against which baseline, and how much supporting evidence can outsiders examine? A headline speedup alone answers none of them.

Publication history

Nov 6, 2026, 21:08 WIB — First published.

No published corrections or earlier versions.

How this was made

UpThat uses AI-assisted research and writing. Source checking and publication follow our editorial review workflow. See the original documents above to assess the reporting.

How we use AI ↗

Technology · news

METR finds faster AI research at Anthropic. The evidence is only partly public.

An independent assessment estimates roughly twice the historical pace on well-defined tasks, while redactions limit outside scrutiny.

By UpThatNov 6, 2026, 21:08 WIB3 min read4 sources

An early warning is a different question from immediate danger

Access for evaluators does not guarantee visibility for readers

Publication history

Nov 6, 2026, 21:08 WIB — First published.

No published corrections or earlier versions.

How this was made

UpThat uses AI-assisted research and writing. Source checking and publication follow our editorial review workflow. See the original documents above to assess the reporting.

How we use AI ↗

Reading size

METR finds faster AI research at Anthropic. The evidence is only partly public.

An early warning is a different question from immediate danger

Access for evaluators does not guarantee visibility for readers

Publication history

How this was made

Only what’s worth bringing up.

Reading size

METR finds faster AI research at Anthropic. The evidence is only partly public.

An early warning is a different question from immediate danger

Access for evaluators does not guarantee visibility for readers

Publication history

How this was made

Only what’s worth bringing up.

PagerDuty's AI SRE Agent Can Investigate Outages. It Still Needs Guardrails to Fix Them.

An early warning is a different question from immediate danger

Access for evaluators does not guarantee visibility for readers

Publication history

How this was made

Next, worth bringing up

AWS Web Answer packages search and synthesis into one agent tool

Claude Fable 5.5: Release Date, Leaks and Verified Facts

Only what’s worth bringing up.

An early warning is a different question from immediate danger

Access for evaluators does not guarantee visibility for readers

Publication history

How this was made

Next, worth bringing up

AWS Web Answer packages search and synthesis into one agent tool

Claude Fable 5.5: Release Date, Leaks and Verified Facts

Only what’s worth bringing up.

PagerDuty's AI SRE Agent Can Investigate Outages. It Still Needs Guardrails to Fix Them.