By asahifile 
On 5 November the nonprofit evaluator METR published an eleven-week audit conducted inside Anthropic, and the number that travelled is 1.9. Between October 2025 and October 2026, the benchmark scores of Anthropic’s models — measured on AECI, the company’s version of the Epoch Capabilities Index — rose at an estimated 1.9 times the rate they had managed between March 2024 and October 2025, with a 95 percent confidence interval of 1.5 to 2.8. METR puts it more plainly: about two years of historical benchmark progress in the last year, roughly the distance from Claude 2.1 to Sonnet 4.5.
The report does not claim a machine that improves itself unsupervised. It claims that AI assistance inside Anthropic’s own research most likely accounts for a majority of the speed-up, and that extra training compute was not the cause. What follows reads the arithmetic rather than the headline.
What the auditors were allowed to see
The engagement ran from 12 August to 29 October with 2.5 full-time staff and produced a roughly 100-page full report. Access was unusual: benchmark scores for unreleased models, compute data by domain, employee interviews. Two constraints travelled with it. Anthropic held unilateral redaction rights, and METR’s own AEF-1 checklist marks two requirements unfulfilled — that methods be detailed enough for independent review, and that the provider not hold authority to redact awkward findings. A compute-allocation study that would have split capabilities work from safety work was in the original contract and was cut midway at Anthropic’s request.
Three multipliers for one acceleration
The same phenomenon wears three numbers. Anthropic’s Haiku 5.5 system card reported 1.7 times; METR says 1.9, partly because newer internal models were included. Wire copy rounds both to “nearly twice”. Then there is the employees’ own estimate, plotted from about 1.2 times for Claude 3.7 Sonnet to about 12 times for an internal model called Model 4 — tenfold across 21 months, which METR fits at 3.9 times per year.
Run that against the measured curve and the two do not meet. Tenfold over 21 months is about 3.7 times a year, close enough to the fit to confirm it; but 3.9 sits above the top of the confidence interval, 2.8. The people doing the work feel roughly twice the acceleration the index records — a gap worth keeping separate: self-reported uplift has no denominator.
Eleven weeks is six points of progress
The audit window is itself a unit of measurement. METR’s fitted slope after October 2025 is 29 AECI points a year against 15 before it. Across the 78 days from 12 August to 29 October, that implies about 6.2 points of movement — 41 percent of an entire pre-2025 year — inside the period the evaluators were watching. The seven days to publication add roughly 0.6 points more. Every figure in the report describes a model already several points behind the one shipping as a reader reaches the paragraph.
The eighty percent the report partly takes back
Attribution is where the arithmetic goes soft, and METR says so. If the total rate is 1.9 times baseline, the excess is 0.9; assigning 80 percent of that excess to AI yields an AI-only multiplier of about 1.72, leaving 0.18 for hiring, data and algorithmic luck. But the limitations section concedes that non-AI factors “could explain all or most of the recent acceleration”, and an appendix notes that Anthropic’s historical AECI growth had run below the wider industry frontier before late 2025 — a low base that flatters any ratio computed against it. Hold the measured slope, 29 against 15. Treat the 80 percent as the guess its authors call uncertain.
Who should not use this number
If you are drafting capability thresholds, the useful figure is not 1.9 but the slope: at 29 points a year, an 18-month rulemaking cycle spans 43.5 AECI points against 22.5 at the old rate. Set review cadence shorter than that, and do not project 1.9 past twelve months — METR would not be surprised by 1.3, or by 3. If you are pricing compute, this says nothing about demand. If you are judging whether AI research is already autonomous, it says assistance, not autonomy.
What to watch
- Whether the next assessment reproduces the 29-point slope or drifts toward the 1.3 end of the interval.
- Whether the capabilities-versus-safety compute split is restored — the cheapest number in the exercise.
- Whether OpenAI and Google DeepMind accept the same eleven-week protocol; one lab’s curve is not the frontier’s.
- Any sign of hyperbolic rather than linear acceleration, which METR says would make its estimate an understatement.
Sources: METR, Independent Assessment of AI R&D Acceleration at Anthropic in October 2026; Reuters; AP via US News. Derived figures — 78 days, 6.2 and 43.5 points, the 1.72 multiplier, the 3.7-times-a-year check — were calculated by huanghuajian.com from the published rates.