ResearchPapers, evidence & method
METR estimates AI progress at Anthropic has run about 1.9 times faster since October 2025, mostly because of Anthropic's own AI use
An 11-week METR assessment estimates Anthropic's AI capabilities have improved about 1.9 times faster since October 2025. It attributes roughly 80% of the gain to the company's own use of AI, while much of the supporting evidence was redacted.
From the The Verge · AI article of November 7, 2026 and cited sources: METR fitted a trend with a breakpoint to Anthropic's internal capability index. [7] [4]
SOURCES[7] METR: Independent Assessment of Anthropic AI R&D… metr.org[4] explainx.ai: METR on Anthropic, claimed vs verified explainx.ai
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 trend with a breakpoint to Anthropic's internal capability index. [7] [4]
- Original source
- The Verge · AI article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 7, 2026 article published
- Covered
- Nov 7 edition
- Evidence
- 9 sources across 9 sites
- 1Why a new report on Anthropic automating AI development is…theverge.com
- 2Unite.AI: METR estimates Anthropic's capability growth has…unite.ai
- 3Outside research group says Anthropic is rapidly…usnews.com
- 4explainx.ai: METR on Anthropic, claimed vs verifiedexplainx.ai
- 5TechRepublic covers METR's assessment of AI research…techrepublic.com
- 6CryptoBriefing reports on METR's analysis of Anthropic's…cryptobriefing.com
Reach: not measured yet
METR fitted a trend with a breakpoint to Anthropic's internal capability index. It found annual gains rising from about 15 points to about 29 after October 2025, a 1.9x speed-up with a 95% interval of 1.5 to 2.8x. METR rules out added training compute as the cause and credits most of the gain to AI that helps human researchers. [7] [4]
The analysis rests on internal data that Anthropic approved for sharing and later partly redacted. That includes the breakdowns by pipeline stage and by compute, data and algorithms. Anthropic also cancelled a planned review of how it splits compute between capabilities and safety work, so the core findings cannot be independently replicated. [7] [1] [2]
Three unreleased models finished training between July and October 2026, and Anthropic's own AI R&D warning threshold may have been crossed around mid-July. Staff using the most capable model reported in informal polls that they worked about 12 times faster, though METR says those answers may not be representative. [8] [1] [2]
Safety researchers and policymakers criticised both the pace and the lack of transparency. One New York legislator urged Congress to ban recursive self-improvement, and outside researchers said the findings track at or ahead of the AI 2027 scenario's timeline. [1]
a fitted capability trend shows faster progress at one leading lab, and compute is ruled out as the driver.
Read the full assessment
Implication: if AI is speeding up AI research, public model releases and regulatory pressure, including calls to ban recursive self-improvement, may both accelerate.
Executive brief
METR, an outside AI-safety evaluator, estimates that AI capabilities at Anthropic have been improving about 1.9 times faster since October 2025, which it describes as roughly two years of progress in one. METR attributes about 80% of that speed-up to Anthropic's own use of AI (METR). METR also says Anthropic redacted most of the evidence METR considered most informative, so readers have to rely on METR's judgment for the central conclusions (The Verge). The report arrives as Anthropic prepares to go public and as its CEO calls for slowing AI development.
What changed and event timeline
Estimated breakpoint
METR's piecewise fit puts the start of the speed-up here. Before it, Anthropic gained about 15 points a year on its capability index; after it, about 29 ().
Three unreleased models finish training
Models with the codenames Model 2, Model 3 and Model 4 completed training in July, August–September and September–October. The Decoder reports that Anthropic's own AI R&D warning threshold may have been crossed around July 15 ().
Report published
METR released a public version of a report that ran about 100 pages before Anthropic's redactions. Anthropic cofounder Sam McCandlish acknowledged "we redacted a lot" (;).
Backlash
Safety researchers and policymakers pushed back. New York Assemblyman Alex Bores urged Congress to "ban RSI" (recursive self-improvement). Anthropic referred press questions to McCandlish's post ().
Capabilities and access
- Models in scope: Unreleased 2022, Model 2, Model 3 and Model 4. None is public.
- Productivity: Staff using the most capable internal model self-reported working about 12 times faster, roughly double what they reported earlier while using Mythos 5 (The Verge).
- METR's access: METR reviewed information Anthropic had pre-approved.
Read the full section
- Models in scope: Unreleased 2022, Model 2, Model 3 and Model 4. These are codenames given to METR; the models have different internal names. None is public. Model 4 scored above 50% on all 15 benchmarks used (METR).
- Productivity: Staff using the most capable internal model self-reported working about 12 times faster, roughly double what they reported earlier while using Mythos 5 (The Verge).
- METR's access: METR reviewed information Anthropic had pre-approved. It did not get unmediated "employee-like" access, and a small number of its data requests were denied (Unite.AI).
Technical analysis for researchers and developers
- Metric: AECI, Anthropic's version of Epoch's capabilities index, which combines benchmark scores through a logistic link function and assumes task difficulty is symmetrically distributed.
- Result: 1.9× faster progress, 95% confidence interval 1.5–2.8×. Alternative modelling choices give 1.1–2.9×.
- Compute ruled out: METR says more training compute did not drive the acceleration.
Read the full section
- Metric: AECI, Anthropic's version of Epoch's capabilities index, which combines benchmark scores through a logistic link function and assumes task difficulty is symmetrically distributed. METR fitted a piecewise linear trend with an estimated breakpoint and compared March 2024–October 2025 against the period after that (METR).
- Result: 1.9× faster progress, 95% confidence interval 1.5–2.8×. Alternative modelling choices give 1.1–2.9×.
- Compute ruled out: METR says more training compute did not drive the acceleration.
- Reproducibility: Low. The breakdowns by pipeline stage, the split between compute, data and algorithms, and the bottleneck analysis were all redacted. Because Model 4 scores above 50% on every benchmark, the benchmarks are approaching saturation, which weakens future fits.
Claims and evidence
- Speed-up of about 1.9×, mostly due to AI. No public data allows a separate replication.
- About 12× productivity. Comes from informal Slack polls of staff.
- This is Anthropic's justification. METR says most of the redacted material was "very high-level" (The Verge).
Read the full section
- Speed-up of about 1.9×, mostly due to AI. Reported by METR, an external evaluator, but based on internal data that Anthropic chose to share and then partly redacted (METR). No public data allows a separate replication.
- About 12× productivity. Comes from informal Slack polls of staff. METR cautions that the answers took little effort and may not represent staff broadly (Unite.AI).
- Redactions protect against rivals. This is Anthropic's justification. METR says most of the redacted material was "very high-level" (The Verge).
- Range of uncertainty. METR says the true figure could plausibly be anywhere from 1.3× to 3× (explainx.ai).
Context and prior work
- AI 2027, the AI Futures Project scenario, placed a key self-improvement breakthrough in early 2027.
- OpenAI said in September that it was making progress toward an automated AI researcher by March 2028.
- The assessment is separate from METR's newer program that gives safety researchers "employee-like" access to Anthropic's systems (The Verge; Wikipedia: METR).
Read the full section
- AI 2027, the AI Futures Project scenario, placed a key self-improvement breakthrough in early 2027. Marius Hobbhahn (CEO of Apollo Research) and Sydney Von Arx (founder of The Nightingale Collective and a former METR researcher) both said the findings track at or ahead of that timeline (The Verge).
- OpenAI said in September that it was making progress toward an automated AI researcher by March 2028.
- The assessment is separate from METR's newer program that gives safety researchers "employee-like" access to Anthropic's systems (The Verge; Wikipedia: METR).
Limitations, safety and contested findings
- Could be an overestimate: non-AI factors may explain more of the speed-up, the gains may be limited to easy-to-grade benchmark tasks, and Anthropic's historical baseline was unusually low compared with the industry (METR).
- Could be an underestimate: growth may be speeding up further, and AECI may understate real-world impact.
- Autonomy: Most of the speed-up comes from AI assisting human researchers.
Read the full section
- Could be an overestimate: non-AI factors may explain more of the speed-up, the gains may be limited to easy-to-grade benchmark tasks, and Anthropic's historical baseline was unusually low compared with the industry (METR).
- Could be an underestimate: growth may be speeding up further, and AECI may understate real-world impact.
- Autonomy: Most of the speed-up comes from AI assisting human researchers. METR sees signs that fully autonomous AI contributions will grow and found no clear bottlenecks.
- Criticism: Daniel Kokotajlo of the AI Futures Project called Anthropic's conduct "recklessly gambling with our lives." Brianna Rosen of the Institute for AI Policy and Strategy said the report was not transparent enough about its evidence (The Verge).
- Unverified chart: The Decoder reports that one graph in the report is labelled "FAKE DATA" (The Decoder).
Business and practitioner implications
- Expect faster releases: three trained but unreleased models suggest Anthropic's next public models could arrive quickly and with large jumps in capability.
- Regulatory risk is rising: Bores' call to ban RSI and the incoming Democratic majority make new rules more likely.
- Read evaluator reports carefully: if the lab controls what gets redacted, an outside assessment is weaker evidence than it looks.
Read the full section
- Expect faster releases: three trained but unreleased models suggest Anthropic's next public models could arrive quickly and with large jumps in capability.
- Regulatory risk is rising: Bores' call to ban RSI and the incoming Democratic majority make new rules more likely. Anthropic's IPO adds pressure for disclosure.
- Read evaluator reports carefully: if the lab controls what gets redacted, an outside assessment is weaker evidence than it looks. Procurement and governance teams should ask what the evaluator could access and what was removed.
- Developers: the bottleneck appears to be moving from writing code to generating ideas and securing compute (The Decoder).
- What's next: Anthropic says it will publish its own report on the pace of AI progress in the coming weeks.
The source trail.
Sources (9)
Why a new report on Anthropic automating AI development is alarming researchers
Article text retrieved; extracted text may omit tables or interactive elements.
theverge.comUnite.AI: METR estimates Anthropic's capability growth has nearly doubled
unite.aiOutside research group says Anthropic is rapidly accelerating AI development amid calls to slow down
Related coverage; assess separately
usnews.comexplainx.ai: METR on Anthropic, claimed vs verified
explainx.aiTechRepublic covers METR's assessment of AI research acceleration at Anthropic
Related coverage; assess separately
techrepublic.comCryptoBriefing reports on METR's analysis of Anthropic's models
Related coverage; assess separately
cryptobriefing.com