AI now builds AI twice as fast: the METR report on Anthropic and what it means for your business
Direct answer
On November 5, 2026, independent evaluator METR published its assessment of Anthropic's internal work: the company is improving its models' capabilities on well-specified tasks at 1.9 times the pace recorded between March 2024 and October 2025, with a 95% confidence interval of 1.5 to 2.8 times. METR estimates that pace equals two years of historical progress in a single year and figures that about 80% of the current acceleration comes from using AI itself in research and development. The report's central warning: Anthropic had the right to redact non-public information and redacted most of the evidence METR found most informative, so the conclusion rests largely on the evaluator's word.
The numbers that matter
- 1.9 times faster since October 2025: the figure comes from AECI, Anthropic's version of the Epoch Capabilities Index, which combines benchmark scores. The 95% confidence interval runs from 1.5 to 2.8 times; under other modeling choices, the range runs from 1.1 to 2.9 times.
- From 15 to 29 points per year: annual AECI growth rose from about 15 points before October 2025 to about 29 points after. METR reads that as the equivalent of the jump from Claude 2.1 to Claude Sonnet 4.5 in one year.
- 11 weeks inside Anthropic: METR's team worked from August 12 to October 29, 2026 with access to internal training data, benchmarks, algorithmic efficiency figures and about 15 staff interviews, at 2.5 full-time equivalents. METR took no payment.
- About 80% attributed to AI: METR estimates most of the current acceleration comes from AI use in research, and confidently rules out faster growth in training compute as the cause. METR itself acknowledges substantial uncertainty in that attribution.
- 3 unreleased models included: the analysis counted three models trained since July that have not launched yet, alongside the public models.
What it means for a business that buys AI
- Improvement cycles are shortening: if the pace holds, the model you choose today can fall behind in months, not years. The vendor decision becomes a periodic review, not a one-time purchase.
- The acceleration has no visible limit: METR says it found no convincing bottlenecks in algorithms, compute or data that would prevent further progress over the next 12 months. Plan as if improvement continues, with contracts that let you switch models.
- The redacted evidence is the honest data point: the report is valuable precisely because it says what it cannot prove. When a vendor sells you a performance number, ask for the methodology and the confidence interval, just as METR published its own.
- Risk accelerates too: METR paired the report with a review of the automated-research section of Anthropic's August 2026 risk report, questioning whether the company had crossed its own 2x acceleration early-warning threshold: it calls that plausible and says the public report understates the acceleration that was occurring.
Note: this article summarizes METR's November 5, 2026 report, “Independent Assessment of AI R&D Acceleration at Anthropic in October 2026,” with coverage from Reuters and the Associated Press, and analysis from ByteHerald, MadRobot, i10x and alfw.ai. The figures are METR estimates on benchmarks of well-specified tasks, not a measure of performance on a business's real work. The 80% attribution to AI is a METR estimate with substantial uncertainty acknowledged by the evaluator itself.
Want AI that works for your business?
Message us on WhatsApp and let's talk about automation and AI agents for your business.
WhatsApp inicIARelated articles
- Claude Haiku 5.5 costs up to 90% less: the AI price war continues
- OpenAI paused its most powerful model for failing its own safety standards
- AI ROI: the real 2026 numbers
Frequently asked questions
What did METR find about Anthropic?
That Anthropic is improving its models' capabilities on well-specified tasks at 1.9 times the pace of March 2024 to October 2025, with a 95% confidence interval of 1.5 to 2.8 times. METR describes that as two years of historical progress in one year.
Why does METR think AI is the cause?
Because it confidently ruled out faster growth in training compute as the cause, and estimates about 80% of the current acceleration comes from AI use in research. METR itself acknowledges substantial uncertainty in that attribution.
How solid is the evidence?
Limited by design: Anthropic redacted most of the evidence METR found most informative. The report publishes confidence intervals and robustness ranges (1.1 to 2.9 times) instead of a single figure.
What should a business do with this information?
Treat model selection as a periodic review, not a one-time purchase, and demand methodology and confidence intervals when a vendor sells a performance figure.
