An outside auditor says Anthropic's AI has doubled its research speed. Anthropic withheld most of the key evidence

Weeks before what could be the largest stock offering ever, Anthropic has a new number to explain. METR, a nonprofit that tests AI systems, says the company's models are now improving about twice as fast on well-defined tasks as they did between March 2024 and October 2025. Its best guess is that roughly 80% of that speed-up comes from Anthropic using its own AI to do research.

That matters well beyond safety circles. A doubling in pace driven by AI doing research is the exact signal Anthropic's own safety rulebook treats as a warning sign, The Decoder notes. For corporate buyers, it suggests vendor models may change faster than procurement cycles assume. For investors, it is a claim about competitive speed they mostly cannot verify.

That is the catch. Anthropic redacted most of the evidence METR considered most informative. METR says the public now has to rely on its judgment "much more than would be ideal."

What the auditor found

About two and a half full-time METR staff worked with Anthropic's internal data from August 12 to October 29, after the engagement was stretched from four weeks to eleven. They reviewed test scores for internal models, training-compute figures, staff polls and roughly 15 interviews. METR says it took no payment, only free AI usage credits.

The central measure is Anthropic's internal version of a composite score, built by research group Epoch AI, that blends many standard AI tests into one number. Since October 2025, that score has climbed 1.9 times faster than before, with a 95% confidence range of 1.5 to 2.8 times. Put simply, about two years of past progress fit into the last year.

METR said it was confident the jump did not come from faster growth in training compute, the raw computing power used to build models. It found no convincing bottlenecks in algorithms, computing power or data over the next 12 months, though it said it was not confident in that view. It warned of "a significant risk" that progress could accelerate further over the coming year.

The source of the speed matters. METR says AI cannot yet make highly significant, fully independent contributions to the most important model-training work. The gain comes mainly from researchers working faster with AI assistants. But Anthropic researchers described "signs of life" of more autonomous AI contributions, METR and the Associated Press reported.

The parts outsiders cannot see

Anthropic held the right to redact anything unilaterally. According to METR, CryptoSlate and The Decoder, it removed forecasts for future models, breakdowns by stage of training, bottleneck analysis and estimates of algorithmic progress. METR's own disclosure checklist says it could not share its methods in enough detail because of the redaction rules.

Anthropic also cancelled, midway through, the part of the review that would have checked how it divides computing power between building capabilities and safety research. It said it wanted to keep the work focused on risk assessment and would revisit the issue. Forecaster Daniel Kokotajlo argued such audits are needed to make any promise to slow down checkable, The Decoder reported.

Co-founder Sam McCandlish, Anthropic's responsible scaling officer, wrote on X that "AI is speeding up our research, but it's hard to say exactly how much," according to the AP. He said detailing what drives the speed-up risks "handing out a road map for others to go faster."

Not everyone accepts that trade-off. Jeffrey Ladish, who runs AI safety group Palisade Research, said he trusts METR, but "I don't think that the world should have to trust METR on this one," BigGo Finance reported. He argued pressure should fall on Anthropic to release the data.

Why the timing is awkward

METR also reviewed the research-speed section of Anthropic's August public risk report and concluded it understated the acceleration, the AP reported. The Decoder adds that METR considered it possible the warning level had been reached as early as July 15, though CryptoSlate reports both groups leaned toward it not having been crossed. Either way, Anthropic's amended IPO filing concedes it may spot such speed-ups only after they happen, TechCrunch reported.

It also sits uneasily beside Anthropic's own messaging. In a September 12 essay, chief executive Dario Amodei urged the industry to slow the pace of capability gains so safety work could keep up, and named METR as a suitable outside evaluator. The AP framed the new findings as rapid acceleration "amid slowdown calls." Beyond McCandlish's post, Anthropic did not immediately comment to the AP.

There are real reasons for caution about the headline figure. METR says gains may be concentrated in tasks with clear right answers, which may not carry over to messy real-world work. Anthropic's earlier baseline may have been unusually slow, inflating the comparison. Run across all frontier developers, the same analysis showed only 1.4 times acceleration, which may partly reflect Anthropic's models having trailed the industry's best in 2025.

Staff surveys are shakier still. Informal polls put self-reported productivity gains at about 12 times, which METR calls very loose evidence. One Anthropic researcher put AI's share of the gains at just 15% to 20%, The Decoder reported.

For decision makers, three things are worth watching before the listing. First, whether Anthropic releases any redacted evidence or restarts the compute audit. Second, whether OpenAI and Google grant outside auditors similar access to their own research pace. Third, whether the scaling-policy updates Anthropic told IPO investors it is considering add hard limits on AI-driven research, or simply more reporting. For now, the most important number about the pace of frontier AI rests largely on trust.