The director’s view
Is AI accelerating its own development? METR examines the evidence
The difficult part is separating faster progress from its causes. A new assessment makes that distinction explicit.

METR’s 5 November assessment estimates that Anthropic’s progress on well-specified AI tasks is roughly twice its March 2024–October 2025 pace. The authors consider 1.3–3× plausible and judge that AI-assisted research likely explains most of the acceleration.
A faster rate is a specific claim
The unit matters. If a fixed capability measure formerly gained ten points in a year and now gains twenty, its progress rate has doubled. That arithmetic says nothing by itself about the time needed to finish a software project or the chance of a correct answer. Those require their own measurements.
METR’s AECI benchmark trend gives 1.9× acceleration, with a 95% confidence interval of 1.5–2.8×. Keep this fitted result separate from the broader judgment above.
Sources: METR [1]
How different tests become a trend
Epoch’s public ECI implementation fits benchmark scores using a curve that relates model capability, test difficulty and how sharply performance changes. It estimates a shared capability scale rather than treating a percentage from one test as interchangeable with a percentage from another. The public code makes that mathematical approach inspectable; it does not supply Anthropic’s private inputs.
Sources: Epoch AI [2]
Inspect the public ECI implementation
Code for fitting a capability index from benchmark scores. It is methodological context, not METR’s private analysis notebook.
The causal claim remains harder to audit
Anthropic redacted most of the evidence METR found especially informative. Readers therefore cannot fully audit the attribution to AI.
Observing a steeper curve and identifying its cause are different research steps. A causal estimate needs a comparison with the progress that would have occurred without the intervention. That alternative history cannot be read directly from the observed curve. The useful reading is therefore two-part: evaluate the reported acceleration, then ask how strongly the available evidence isolates AI’s contribution. Neither number alone gives a date for autonomous AI research.
Sources: METR [1]
SOURCES & METHOD
The record behind the article
Analysis of the published assessment and the public index methodology. We did not inspect Anthropic’s confidential evidence or reproduce the analysis.
