The METR Anthropic AI report, published November 5, 2026, makes a claim that would be striking even coming from the company itself: Anthropic's models are improving about twice as fast as they did from March 2024 to October 2025, and AI tools drive most of that speedup. The claim comes from METR, a nonprofit that evaluates risks from advanced AI. It arrives with a serious catch. METR says Anthropic redacted most of the evidence it found most informative.
What the METR Anthropic AI report found
METR's headline estimate is that, as of October 2026, Anthropic is likely improving its models on well-specified tasks at about twice the rate of its earlier baseline. The estimate rests on benchmark scores. According to METR's full assessment, Anthropic's models have improved about 1.9 times faster since October 2025, with a 95% confidence interval of 1.5 to 2.8 times.
Those numbers carry a lot of uncertainty. METR puts its high-uncertainty range for the overall rate at roughly 1.3 to 3 times. A figure that could mean anything from a modest bump to a tripling is a warning sign, not a forecast.
METR, short for Model Evaluation and Threat Research, was founded in 2022, and its CEO and founder is Beth Barnes. A large share of its work tests whether AI systems can perform AI research, including through a benchmark called RE-Bench, according to Wikipedia. That focus is why this report lands with such weight.
What does twice as fast mean in practice? If the pace holds, a stretch of progress that once took a year would take about six months. That compression is the reason the finding has drawn attention beyond research circles.
The second finding concerns cause. METR estimates that perhaps 80% of the current acceleration comes from AI being used in Anthropic's research and development work. METR presents that number as a guess rather than a measurement. It also says the late-2025 speedup was likely influenced by factors that had nothing to do with AI.
METR also says it is confident the acceleration is not explained by faster growth in training compute. That matters because the simplest alternative explanation is spending. Bigger training runs can lift scores without any change in how research gets done, and METR is arguing that the pace reflects something about the research process itself.
How METR measured the speedup
METR's fieldwork ran from August 12 to October 29, 2026, about 11 weeks. The team had access to internal data and interviewed staff. It delivered a full report of roughly 100 pages to Anthropic on October 22, and the public version followed after several rounds of redaction review.
Staff views on the speedup are mixed. Technology.org reports that, among about 15 staff interviewed, the views broke down this way:
- Two researchers clearly backed a speedup of at least two times.
- Three leaned that way.
- Four were unclear.
- Two disagreed.
Staff Slack self-reports point in the same general direction, though METR is cautious about them. Staff rated the uplift from the newest internal model at about 12 times on average in October. METR called that "only very loose evidence," as WebProNews quoted it.
The result also depends on which models are counted. When METR added other labs' models to the trend, the acceleration fell to about 1.4 times. Commentators in the AI forecasting community have said the 2x figure roughly matches their expectations, but the frontier-wide number is considerably lower. That gap is one of the most useful things a reader can take from the report.
The findings also bear on a claim Anthropic has made about itself. The company has warned that its models may speed up their own development, a warning USA 24 has covered. The Decoder reports that METR believes Anthropic may have crossed its own AI-acceleration warning threshold back in July.
Why most of the evidence is hidden
The redactions are the central problem with this report. METR wrote in a November 5 post on X: "For the public version, Anthropic redacted a majority of the evidence we found most informative." According to the reporting, the removed material includes per-stage acceleration estimates and bottleneck analysis.
Anthropic's stated reason is that disclosure could help rivals. METR says the cuts removed much of the most informative evidence and limited outside scrutiny. Anthropic's full response was not available in the coverage reviewed for this story, so readers should treat the company's rationale as stated, not tested.
That gap has a practical consequence. Outsiders cannot independently verify the 1.9 times figure or the 80% attribution. They can read METR's reasoning and its stated uncertainty, but they cannot check the underlying per-stage numbers that would show which parts of the research pipeline sped up.
Two other threads add to the picture. A companion METR review examines Anthropic's redacted August 2026 risk report, which was published as a PDF. A planned compute-allocation audit was also reportedly pulled in the middle of its pilot. The reporting gives limited detail on why, and it is one of the open questions most worth tracking.
Where the METR Anthropic AI report is weakest
METR is unusually candid about its limits, and readers should weigh the findings with those limits in view.
- METR says its view changed several times during the study.
- It saw no significant bottlenecks in algorithms, compute or data over the next 12 months, but says it is not confident in that assessment.
- The measured speedup relies on verifiable tasks, which METR acknowledges may overstate the effect.
- The frontier-wide estimate of about 1.4 times sits well below the 1.9 times Anthropic-specific figure.
Taken together, those caveats argue for treating the report as a serious signal rather than a settled measurement. The most useful next question is whether the same method can be applied to other labs, and whether future evaluations include the evidence that is currently hidden.
What the findings mean for policymakers, investors and developers
Policymakers have been pressing for federal AI oversight, and USA 24 has reported on efforts to list federal laws that bind AI labs. None of the sources reviewed tie this report to a specific legislative action. Its value for regulators is still concrete, because it shows what a verification gap looks like when a lab's own disclosures cannot be checked from outside.
Investors and enterprise buyers face a similar problem. Anthropic's planned public offering is drawing scrutiny over its valuation, a debate USA 24 has followed. A third-party estimate of the company's research pace now exists, but the company's public disclosures did not fully allow outsiders to verify it. That makes the estimate useful for asking questions, but not yet a solid basis for confidence.
Developers building on Claude models have the most direct stake. If METR's estimate holds, model iteration could speed up and upgrade cycles could shorten. Teams should plan for more frequent model changes, keep evaluation suites current, and avoid hard-coding behavior that a new release could change.
Practical steps for readers
- Read METR's public report and its redaction notice directly rather than relying on headlines.
- Treat the 80% attribution as an estimate, and treat the 1.9 times figure as a range with wide bounds.
- Compare the Anthropic-specific benchmark result with the 1.4 times frontier-wide figure before citing either one.
- Watch for Anthropic's formal response and for any follow-up audit that covers the redacted sections.
What comes next
The next test is whether the finding holds up without relying on what Anthropic chooses to publish. Three developments would show how much weight it can bear: whether the compute-allocation audit resumes and what it finds, whether Anthropic responds to the warning-threshold question, and whether METR or other evaluators run comparable assessments on rival labs. If similar acceleration shows up elsewhere, the story moves from one company's disclosure to a broader pattern in frontier AI development. If it does not, the 2x claim will look more like a feature of one lab's workflow.
