Anthropic Is Improving Its AI Twice as Fast as Before, Its Own AI Is Doing Most of the Speeding Up, and Most of the Proof Is Redacted
Two years of progress in one year
Progress in AI used to have a simple story: bigger training runs, more compute, better models. If that story were still the whole story, the pace of progress would track spending on chips. An outside auditor has now spent 11 weeks inside one frontier lab and concluded that something else is doing the work.
On November 5, METR, the nonprofit evaluator formally known as Model Evaluation and Threat Research, published its independent assessment of AI research and development at Anthropic. Its central finding: as of October 2026, Anthropic is likely improving the capabilities of its models on well-specified tasks at around twice the rate it managed between March 2024 and October 2025, and a majority of that acceleration is likely due to Anthropic using its own AI to do the research. That is not recursive self-improvement in the science fiction sense, and METR is careful to say so. It is something more concrete: the lab’s main tool for building the next model has become the current model.
What METR actually measured
METR staff, 2.5 full-time equivalents, worked at Anthropic from August 12 to October 29, 2026, with access METR calls unprecedented: internal benchmark scores for released and unreleased models, pretraining and reinforcement learning data, training compute figures, speedup analyses, forecasts, and staff interviews. METR took no payment, and each side covered its own costs.
The headline number comes from AECI, Anthropic’s version of the Epoch Capabilities Index, which combines benchmark results into a single capability score:
- Before October 2025, AECI was rising at about 15 points a year.
- Since October 2025, it has risen at about 29 points a year.
- The ratio is an estimated 1.9 times acceleration, with a 95 percent confidence interval of 1.5 to 2.8 times.
- In plain terms, METR says Anthropic made roughly two years of historical progress in the last year, about the distance from Claude 2.1 to Sonnet 4.5.
Two checks keep the number honest. Anthropic’s own Claude Haiku 5.5 system card reported 1.7 times; METR’s higher estimate mainly reflects three newer unreleased models trained since July. When METR varied which benchmarks and models to include, estimates ranged from 1.1 to 2.9 times. The direction survived every variation, the exact size did not, and METR would not be surprised if the true figure were closer to 1.3 or to 3 times. Treat 1.9 as the centre of a range, not a precise measurement.
It was not the compute
The most consequential sentence in the report may be the shortest one. METR says it is confident the acceleration was not due to an acceleration in the scaling of training compute. The detailed evidence for that was redacted, but the conclusion was published plainly.
That separates two futures. If progress were still mainly bought with compute, it could be forecast from data centre announcements. If it is now mainly produced by researchers working much faster with AI help, the constraint is something outsiders cannot see: how good the lab’s internal tools and workflows have become.
On attribution, METR offers a guess and labels it as one. It estimates that a majority of the current acceleration, perhaps about 80 percent, is due to AI, while cautioning that other factors likely played a large role in the initial speed-up in late 2025 and that Anthropic’s own staff disagreed with each other. Interview estimates centred around 2 times, but ranged from enthusiasts to a sceptic who credited AI with only 15 to 20 percent. Informal Slack polls in which staff voted on their speed-up from the latest model reached 12 times in October 2026. METR calls that only very loose evidence, and it is right to: individual speed-ups are not the same thing as a faster rate of capability progress.
Assistance, not autonomy, with signs of life
METR draws a line that much commentary will blur. Today, AI accelerates Anthropic’s research primarily by assisting people: writing code, running experiments, and letting researchers try ten ideas and see what sticks, in one interviewee’s phrase. AI is not yet able, METR says, to make highly significant and fully autonomous contributions to the most important parts of training. But it reports significant signs of life for autonomous contributions arriving soon, with much of that evidence redacted. Assistance is the present, autonomy is the direction, and nobody outside the lab can see how close the second step is.
METR also looked for the brakes and did not find them. It saw no clear bottleneck in algorithmic progress, training compute, experiment compute, or data that would prevent further progress in the next 12 months, while admitting its understanding changed several times during the assessment. Its conclusion follows: there is a significant risk that further progress leads to even faster acceleration, because each better model is also a better research tool.
The evidence problem: most of it is blacked out
Anthropic held the right to redact any non-public information, and used it. METR states that Anthropic redacted a majority of the evidence it found most informative for its high-level risk conclusions, covering algorithmic progress estimates, forecasts for future models, and breakdowns by training stage. Anthropic described the redactions as aimed at removing confidential information that could accelerate other AI developers. METR’s verdict is unusually blunt for an auditor: the report does not achieve transparency about evidence, and the public has to rely on METR’s judgment much more than would be ideal, given the weight of the questions answered here.
Two process details belong next to that warning. METR received no sample access to the non-public models behind the numbers, so it could not run its own tests on them. And a planned check on how Anthropic allocates compute between capabilities and safety work was dropped after Anthropic changed the scope mid-assessment, to be revisited later. None of this makes the finding false. METR is independent, unpaid by Anthropic, and visibly willing to publish conclusions its host would rather soften. It does mean the finding is an expert judgment based on evidence you cannot inspect, and it should be quoted that way.
An early warning that may already have gone off
METR published a companion review the same day, of the automated R&D section of Anthropic’s August 2026 Risk Report, and it is the more uncomfortable document. It agrees that risk from models as of July 15, 2026 was low. But it says Anthropic’s claim not to have crossed the 2x acceleration threshold, an early warning line for exactly this dynamic, rests on contestable judgment calls, that it is plausible the threshold had been crossed, and that the public report overall understates the level of acceleration that was occurring at the time.
Read together, the timeline sharpens: by mid-2026 the lab may already have been at its own warning line, and by October its auditor’s best estimate sat right at that line, with an interval running well above it. Anthropic’s chief executive, Dario Amodei, warned in a September essay that since roughly the summer, AI has been advancing drastically faster, driven primarily by AI’s growing ability to build the next generation of AI. METR’s report is the first outside, data-backed check of that claim, and it broadly confirms it.
What to do with this, this week
- Stop forecasting model progress from compute alone. If the auditor is right, capability jumps now depend partly on internal AI-assisted research speed, which no capex chart shows. Shorten your planning horizon for what supplier models will be able to do.
- Measure your own engineers’ speed-up properly. Do not copy the Slack-poll method. Time real tasks, with and without AI help, on your own work. METR’s report is a lesson in how wide self-reported numbers swing.
- Quote the range, not the headline. Write 1.9 times with a 1.5 to 2.8 interval and a wider 1.3 to 3 judgment range, note that the underlying evidence is redacted, and note that the claim covers well-specified tasks, not all work.
- Ask vendors about their internal acceleration, not just their model. If AI-assisted R&D drives progress, a lab’s tooling and research workflow are part of its roadmap, and roadmaps built on last year’s pace will be wrong in one direction: too slow.
- Treat safety timelines as moving with the same multiplier. Evaluation and governance processes designed for the old pace inherit the acceleration too. If capabilities can move at twice the rate, review cycles have to be redesigned, not just hurried.
The bottom line
The strongest evidence yet that AI is speeding up its own development comes with an asterisk the size of the report itself: the detailed proof is redacted, the key figure is an estimate with a wide range, and the mechanism so far is assisted researchers, not autonomous ones. Even with every asterisk attached, the direction is hard to argue with. An unpaid outside auditor, given 11 weeks inside the lab, concluded that progress doubled, that compute did not cause it, and that AI itself likely caused most of it. The race is no longer only about who can buy the most compute. It is about whose AI is best at building the next one, and that contest compounds.