Anthropic's AI is speeding up its own research, METR finds

Anthropic's models are improving roughly twice as fast as they were a year ago, and the company's own researchers credit AI for much of the lift. That is the headline finding of an 11-week, high-access evaluation by the AI safety nonprofit METR, whose staff were embedded at Anthropic from August 12 to October 29, 2026. The report landed Thursday and is now alarming safety researchers, largely because Anthropic has spent a year arguing labs should slow down.

What METR measured

METR tracked Anthropic's benchmark progress using AECI, the company's internal version of the Epoch Capabilities Index, and put acceleration at 1.9x since October 2025, with a 95% confidence interval of 1.5x to 2.8x. In plain terms, Anthropic made as much benchmark progress in twelve months as it did between October 2023 and October 2025.

Crucially, METR says it is "confident that the acceleration was not due to an acceleration in the scaling of training compute." The majority of the speedup as of October, the researchers write, is attributable to AI, and that share is growing. Employee interviews back the numbers: staff using Anthropic's most capable internal model self-reported being around 12 times more productive with AI, roughly double what they reported months earlier on Mythos 5.

Three trained models nobody has seen

Part of the gap between public and internal progress comes from models Anthropic has finished but not shipped. METR's inventory lists Model 2, finished in July 2026, Model 3 in August to September, and Model 4 in September to October. Model 4 scores above 50% on all fifteen benchmarks METR used to position it, a saturation level that leaves the researchers "somewhat nervous" about measuring future progress at all.

The redaction problem

Anthropic had the contractual right to redact the public version of the report, and METR says it used that right to remove "a majority of the evidence we found most informative." Cut from public view: intermediate capability measurements, algorithmic progress estimates, bottleneck forecasts, and most of the breakdown of where the gains came from. Anthropic framed the redactions as removing material that "could accelerate other AI developers." METR's own conclusion is blunter: the public now has to trust the auditors' judgment far more than transparency would ideally require.

Why it matters

The finding pushes recursive self-improvement from thought experiment into quarterly reporting. Much of today's acceleration still runs through humans in the loop, but METR flags "some signs that fully autonomous contributions will play a larger role soon," and reports no clear bottleneck in sight. For Anthropic, heading toward a public listing while its CEO calls for pacing the frontier, the optics are awkward. For everyone building on these models, the takeaway is that capability jumps may arrive faster and with less external warning than release schedules suggest.

Anthropic co-founder Sam McCandlish said the company plans to publish its own analysis of how fast progress is moving inside the lab.