METR finds Anthropic's AI R&D accelerating at a sustained 2x since late 2025, mostly AI-driven but with very uncertain evidence.
Parv Mahajan says Anthropic's August 2026 risk report overstates confidence, and METR would likely conclude the 2x warning threshold has been crossed.
Mahajan urges Anthropic to release raw data on human versus AI contributions so the public can analyze it.
METR finds Anthropic near its own 2x warning line
METR got unusual access inside Anthropic to ask whether the company's AI research is speeding itself up, and found roughly 2x acceleration against historical trends since late 2025, mostly from AI. Anthropic's public risk report presents its 2x warning threshold as uncrossed, yet the company's own numbers land at about 1.97x. The published evidence is mostly well-specified benchmark tasks, so outsiders cannot fully check the work, and METR's methods cannot yet confirm or exclude an intelligence explosion.
Takeaways
METR measures roughly 2x acceleration at Anthropic. METR's report found a sustained 2x acceleration against historical trends since late 2025, or two years of AI progress in one. Mahajan says it is mostly due to AI, though the evidence is uncertain .
Anthropic's own numbers sit near its 2x line. Anthropic's August 2026 risk report said it had not crossed its acceleration warning threshold, yet its own numbers give about 1.97x, close to the 2x line. Mahajan says METR with full evidence would likely say the threshold was crossed .
No bottlenecks found, but confidence stays low. METR found no observed bottlenecks in algorithms, training compute, experiment compute or data over the next twelve months, and no obvious strong diminishing returns. It could not rule out fast acceleration or rule in very fast acceleration. Confidence is low .
Published evidence covers only well-specified tasks. The clearest evidence METR could publish with few redactions was Anthropic's Epoch Capabilities Index, which covers well-specified tasks. Public evidence is compatible with little acceleration on fuzzy, long-horizon work, though METR sees signs of life in autonomous algorithmic improvements .
Methods cannot confirm or exclude hyperbolic growth. METR found no evidence of hyperbolic growth, but its methods could not confidently detect it even if it existed. There is only about a year of data and a limited number of internal models .
Self-reported uplift is close to uninterpretable. Anthropic polled staff on Slack about how much AI sped up their work, and Mahajan warns such studies are extremely hard to interpret . METR found a similar result in 2025, where people reported around 20% uplift but were in fact downlifted .
Redactions leave the public trusting METR. Mahajan wants Anthropic to release raw data, especially on human versus AI contributions to the training pipeline. No single organization should be the arbiter of risk, and redactions leave the public trusting METR more than is ideal .
How METR built its case
Interviews with Anthropic researchers about how their work is changing, and staff self-reports of how much AI speeds them up .
Investigations of different parts of the training pipeline, including how much of it is done by AI versus humans .
Benchmarks of Anthropic's internal models, which METR could run because of the access granted .
A review of the automated AI R&D risk section of Anthropic's August 2026 risk report, published alongside METR's own .
