EcosystemsRegions, accelerators & hubs
METR estimates Anthropic's AI progress has roughly doubled in pace, and AI 2027 authors move superhuman-researcher forecasts into 2027
After 11 weeks embedded at Anthropic, METR estimates the company's internal benchmark gains are arriving about twice as fast as before and attributes most of the speed-up to Claude helping build Claude. Days later, the AI 2027 forecasters moved their timelines into 2027.
From the The Information article of November 11, 2026 and cited sources: METR fitted a piecewise regression to AECI, an index of Anthropic's internal benchmarks, and found a break around October 2025. [2] [7]
SOURCES[2] METR: Independent Assessment of AI R&D Acceleration at… metr.org[7] Washington Post washingtonpost.com
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- METR fitted a piecewise regression to AECI, an index of Anthropic's internal benchmarks, and found a break around October 2025. [2] [7]
- Original source
- The Information article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 11, 2026 article published
- Covered
- Nov 12 edition
- Evidence
- 13 sources across 11 sites
- 1Anthropic, Others Warn of AI Accelerationtheinformation.com
- 2METR: Independent Assessment of AI R&D Acceleration at…metr.org
- 3Crypto Briefing: AI 2027 authors updatecryptobriefing.com
- 4Anthropic: Measuring the pace of AI developmentanthropic.com
- 5TechXplore/APtechxplore.com
- 6METR: Claude Opus 5.5 evaluationmetr.org
Reach: not measured yet
METR fitted a piecewise regression to AECI, an index of Anthropic's internal benchmarks, and found a break around October 2025. Progress since then is about 1.9x faster (95% CI 1.5–2.8x), and other curve choices give 1.1x–2.9x. METR ruled out growth in training compute as the cause. It credits roughly 80% of the speed-up to AI, but says that share is highly uncertain, and Anthropic staff disagreed about it. [2] [7]
The gains mostly come from AI helping human researchers work faster, not from AI running research on its own. They show up on well-specified tasks that can be graded automatically. METR also sees early signs that AI could soon contribute to training by itself. In informal Slack polls, staff said AI sped up their work about 1.2x in January 2025 and about 12x in October 2026. [2] [5] [7]
Anthropic redacted the evidence METR found most informative, including stage-by-stage acceleration figures, bottleneck analyses and forecasts. Co-founder Sam McCandlish defended the redactions as protecting commercially sensitive information. Critic Paul Shearer argues that benchmark trends do not show real-world capability, and that the redacted material is cited only in the alarming direction. [2] [12] [13]
Daniel Kokotajlo of the AI Futures Project moved his median forecast for a superhuman AI researcher from December 2028 to May 2027. The group's six forecasters now put their medians between May and November 2027, and the group called for a pause. The authors call the update preliminary and have acknowledged modeling errors. In September, Anthropic CEO Dario Amodei also urged slowing the pace of AI capability gains. [3] [8] [5]
an outside evaluator with deep internal access measured faster benchmark progress at a frontier lab. Implications: if that pace holds, planning cycles for model-dependent products shorten.
Read the full assessment
Calls to slow down raise regulatory and disclosure pressure, and embedded audits with redactions may become standard.
Executive brief
An outside evaluator, METR, embedded with Anthropic for 11 weeks and concluded that the company's measured capability gains are now arriving about twice as fast as in March 2024–October 2025. METR attributes roughly 80% of that speed-up to Claude being used to build Claude (METR). Days later, the AI Futures Project, the group behind "AI 2027", moved its forecasts for a superhuman AI researcher into 2027 and called for a pause (Crypto Briefing). These are the "two recent reports" behind The Information's column (The Information). Anthropic redacted most of the evidence METR found most informative.
What changed and event timeline
Anthropic publishes pace metrics
Anthropic says Claude "leads" 26% of its AI R&D work, up from under 1% in February. It also reports monitoring about 30,000 research agents ().
Amodei urges slowing down
CEO Dario Amodei called for slowing the pace of AI capability improvement ().
METR evaluates Claude Opus 5.5
METR made a preliminary estimate of about 1.5x AI-driven capability acceleration. It judged the model's gains incremental and unlikely to fully automate AI R&D ().
METR acceleration report
After high-access work from Aug 12 to Oct 29, METR estimated 1.9x acceleration in Anthropic's internal benchmark index. It found no major bottleneck expected within 12 months (;).
AI 2027 authors update
Daniel Kokotajlo moved his median for a superhuman AI researcher from December 2028 to May 2027. The group's six forecasters now give medians between May and November 2027 (;).
Capabilities and access
- METR's report names no model version. Its Opus 5.5 evaluation is the latest model-specific assessment in the reviewed sources (METR).
- The measured gains are on well-specified, automatically gradable tasks (METR).
- The productivity gain comes mainly from AI helping human researchers work faster, not from AI running the lab on its own.
Read the full section
- METR's report names no model version. Its Opus 5.5 evaluation is the latest model-specific assessment in the reviewed sources (METR).
- The measured gains are on well-specified, automatically gradable tasks (METR).
- The productivity gain comes mainly from AI helping human researchers work faster, not from AI running the lab on its own. METR still reports "signs of life" that AI could soon contribute to training autonomously (TechXplore/AP).
Technical analysis for researchers and developers
- Metric: AECI, an index of Anthropic's internal benchmarks. The result was 1.9x acceleration (95% CI 1.5–2.8x).
- Attribution: METR says growth in training compute was ruled out as the cause.
- Uplift data: In Slack polls, staff reported that AI sped up their work about 1.2x in January 2025 and about 12x in October 2026.
Read the full section
- Metric: AECI, an index of Anthropic's internal benchmarks. METR fit a piecewise linear regression with a breakpoint it detected at Oct 7, 2025. The result was 1.9x acceleration (95% CI 1.5–2.8x). Using different link functions (the curve shapes fitted to the scores) gives 1.1x–2.9x (METR).
- Attribution: METR says growth in training compute was ruled out as the cause. The roughly 80% AI share carries substantial uncertainty, and Anthropic staff did not agree on it.
- Uplift data: In Slack polls, staff reported that AI sped up their work about 1.2x in January 2025 and about 12x in October 2026.
- Reproducibility: Low. Stage-by-stage acceleration figures, algorithmic-progress estimates, bottleneck analyses and forecasts were redacted.
Claims and evidence
Read the full section
- Capability pace about 2x. This is METR's estimate, but it rests largely on data Anthropic selected and partly redacted. METR's own range is 1.3x–3x (METR on X).
- Claude leads 26% of R&D work. Anthropic's own figure, by its own definition, not independently checked (Anthropic).
- 12x self-reported speed-up. Based on informal staff polls (METR).
- Superhuman AI researcher in 2027. A forecast, not a measurement. The authors call it preliminary and have acknowledged modeling errors (Crypto Briefing).
Context and prior work
- In May 2026, METR reviewed the automated-R&D section of Anthropic's February 2026 risk report (METR).
- METR's Opus 5.5 evaluation relied on 10 business days of API access, and Anthropic had review rights over the text (METR).
- Coverage of the AI 2027 update cites METR time-horizon doublings every 3–4 months, against the 7 months the original scenario assumed (Better Stack).
Read the full section
- In May 2026, METR reviewed the automated-R&D section of Anthropic's February 2026 risk report (METR).
- METR's Opus 5.5 evaluation relied on 10 business days of API access, and Anthropic had review rights over the text (METR). The November report used much deeper internal access.
- Coverage of the AI 2027 update cites METR time-horizon doublings every 3–4 months, against the 7 months the original scenario assumed (Better Stack).
Limitations, safety and contested findings
- METR's own caveats: gains may be limited to well-specified tasks, and the speed-up could look larger if Anthropic's earlier progress was understated (Washington Post).
- Anthropic's redactions: co-founder Sam McCandlish said the redactions limit transparency but defended them as commercially sensitive (CryptoSlate).
- Independent critique: Paul Shearer argues that benchmark trends and real-world capability diverge, that the uplift evidence is loose, and that redacted material is cited only in the alarming direction (Shearer).
Read the full section
- METR's own caveats: gains may be limited to well-specified tasks, and the speed-up could look larger if Anthropic's earlier progress was understated (Washington Post). Its main safety worry is that AI improving AI could compound.
- Anthropic's redactions: co-founder Sam McCandlish said the redactions limit transparency but defended them as commercially sensitive (CryptoSlate).
- Independent critique: Paul Shearer argues that benchmark trends and real-world capability diverge, that the uplift evidence is loose, and that redacted material is cited only in the alarming direction (Shearer).
Business and practitioner implications
- Timelines: if the 2x pace holds, planning cycles for model-dependent products shorten. Expect capability jumps at more frequent intervals.
- Policy pressure: both Anthropic's CEO and the AI Futures Project are calling for slowdowns or a pause.
- Third-party audits: Anthropic says it will embed outside evaluators (Anthropic).
Read the full section
- Timelines: if the 2x pace holds, planning cycles for model-dependent products shorten. Expect capability jumps at more frequent intervals.
- Policy pressure: both Anthropic's CEO and the AI Futures Project are calling for slowdowns or a pause. That raises regulatory and disclosure risk for frontier labs.
- Third-party audits: Anthropic says it will embed outside evaluators (Anthropic). Embedded evaluation with redactions may become the template for audits, so buyers should read redaction scope as closely as the headline numbers.
- Deal figures in the same newsletter: The Information's excerpt also reports that Stripe acquired OpenRouter (over $7B) and SpaceX acquired Cursor ($60B), and that Perplexity is raising at over $30B. None of these figures is corroborated in the reviewed sources.
The source trail.
Sources (13)
Anthropic, Others Warn of AI Acceleration
Feed excerpt only.
theinformation.com