METR Finds Anthropic AI Acceleration Near Twice Prior Pace

METR says Anthropic is likely improving its AI systems on well-specified tasks at about twice the rate it did from March 2024 through October 2025, according to a public assessment dated Nov. 5, 2026. The independent review covered Aug. 12 to Oct. 29, 2026 and puts the true acceleration somewhere around 1.3x to 3x, with high uncertainty.

What did METR measure at Anthropic?

METR's high-access evaluation asked how much AI is speeding up Anthropic's own R&D. Staff reviewed internal pretraining and reinforcement learning metrics, training compute by domain, internal model benchmarks, AI speedup analyses, and interviews over roughly 11 weeks. The unpaid engagement is separate from Anthropic's Oct. 15 update on independent evaluators.

How large is Anthropic AI acceleration?

On Anthropic's Epoch Capabilities Index (AECI), METR estimates about 1.9x faster benchmark growth since October 2025, with a 95% confidence interval of 1.5x to 2.8x. In METR's framing, that is roughly two years of historical progress packed into the past year. The group says it is confident the jump is not from faster training-compute scaling.

How much of the speedup comes from AI?

METR's subjective guess is that a majority, perhaps about 80%, of the current October 2026 acceleration comes from AI use in R&D. It stresses substantial uncertainty and says late-2025 acceleration likely also had non-AI factors whose details were redacted. Assistance, not fully autonomous research agents, is still the main mode, though METR notes "signs of life" that autonomy may grow soon. Staff Slack-poll self-reports of uplift for the latest internal model reached about 12x in October 2026, which METR treats as loose evidence only.

AP coverage framed the same findings as Anthropic developing AI about twice as quickly as a year ago, even as CEO Dario Amodei has publicly urged the industry to slow the pace of improvement.

Why do the redactions matter?

Anthropic redacted a majority of the most informative evidence from the public report, describing the cuts as removing confidential information that could accelerate other AI developers. METR says that leaves the public leaning more on its judgment than on the full underlying evidence. The same day, METR also published a separate review of Anthropic's Risk Report automated R&D section.

Our Take

This report is distinct from Anthropic's recent recursive self-improvement language in its S-1. METR is measuring something more concrete: an observed ~2x pace on well-specified tasks, mostly attributed to AI helping R&D, with no convincing near-term bottlenecks spotted in algorithms, training compute, experiment compute, or data.

If that loop compounds, capability timelines compress faster than public benchmarks alone suggest. Readers tracking Anthropic's public markets path should also weigh today's separate IPO valuation pushback.

For the financial backdrop, see Anthropic's latest revenue and loss figures.

On the product side, Claude continues to expand into consumer workflows such as Claude Personal Finance.

What are the key METR numbers?

Item	METR finding
Assessment window	Aug. 12 to Oct. 29, 2026
Capability pace vs Mar. 2024 to Oct. 2025	~2x (uncertainty ~1.3x to 3x)
AECI growth since Oct. 2025	~1.9x (95% CI 1.5x to 2.8x)
Share of current acceleration from AI R&D use	~80% subjective guess
Training-compute scaling as cause	Ruled out