Independent evaluators at METR say Anthropic is improving model capabilities on well-specified tasks about twice as fast as it did a year earlier—and that AI itself likely explains most of the current speedup. The public report is one of the clearest looks yet inside a frontier lab’s R&D pace. It is also heavily redacted, narrowly scoped, and timed against Anthropic’s own call to slow down.
On November 5, 2026, the nonprofit Model Evaluation and Threat Research published an independent assessment of AI research-and-development acceleration at Anthropic. After eleven weeks of high-access review from August 12 through October 29, METR concluded that as of October 2026 Anthropic is “likely improving the capabilities of their models on well-specified tasks at around twice the rate they did between March 2024 and October 2025.”
That sentence will travel. It should. It is also not the whole story. METR’s own caveats—uncertainty bands, task scope, redacted evidence, and the difference between AI assistance and autonomous research—matter as much as the headline multiplier. Read without them, the report becomes either a recursive-self-improvement scare story or a proof that safety talk is theater. Neither reading is what the document supports.
What METR measured
METR’s core empirical claim rests on AECI, Anthropic’s version of the Epoch Capabilities Index: an aggregate of internal and public benchmarks that tries to track latent model strength over time. Since October 2025, AECI scores for Anthropic models have grown at an estimated 1.9× the pre-October 2025 rate, with a 95% confidence interval of 1.5× to 2.8×. In METR’s framing, that means roughly two years of historical progress compressed into one year.
The evaluators used ordinary least-squares fits with a single learned breakpoint (around October 7, 2025 in the main specification) and included unreleased internal models that received substantial employee use. Including those internal snapshots is why METR’s 1.9× sits above the 1.7× figure previously reported in Anthropic’s Haiku 5.5 system card.
Robustness checks mostly keep the story near “about 2×.” Changing which snapshots or benchmarks are included typically yields 1.9×–2.2×. Changing the statistical “link function” that maps latent ability to benchmark scores is more consequential: estimates range from 1.1× to 2.9× depending on assumptions about easy versus hard tasks. METR sticks with the logistic default used by Epoch and Anthropic, while noting that alternative fits are almost as good.
One conclusion METR states with unusual confidence: the acceleration was not caused by faster growth in training-compute scale. Details of that evidence were redacted. Everything else about attribution is softer.
AI assistance, not autonomous labs—yet
METR’s subjective judgment is that a majority—perhaps around 80%—of the current (October 2026) acceleration is due to Anthropic’s use of AI for R&D. That guess comes with “substantial uncertainty.” Late-2025 acceleration, by contrast, was likely influenced by non-AI factors whose details Anthropic also redacted.
Staff interviews and surveys are consistent with large AI uplift, but they are not decisive. Slack polls of self-reported speedup reached about 12× for the most capable internal model in October 2026; METR calls that “very loose evidence.” Structured interviews produced a mixed but leaning-positive picture on whether AI was delivering more than 2× acceleration in a researcher’s own area. Some researchers described experiment overhead collapsing; others pointed to hiring or non-AI factors as larger.
Crucially, METR says AI is “not yet able to make highly significant and fully autonomous contributions to the most important parts of the model training process.” Researchers described “signs of life” that such contributions could arrive soon, with maturity varying by pipeline stage. That distinction matters. Assistance that makes humans try ten ideas instead of one is a productivity shock. Fully autonomous improvement of the training stack is a different regime—the one that raises compounding, hard-to-steer risk.
What “well-specified tasks” quietly excludes
METR repeatedly limits the capability claim to well-specified, often automatically gradable tasks. AECI is a benchmark index. It can accelerate while messy, open-ended, real-world impact does not—or the reverse, if impact jumps when models cross thresholds that scores only track smoothly.
The report lists both directions of error. Acceleration could be understated if growth is becoming hyperbolic, if automation is nearer than interviews suggest, or if AECI understates real-world effects (METR cites sudden jumps in math and cybersecurity impact as examples). Acceleration could be overstated if non-AI factors dominate, if models are overfitting to clean tasks relative to messy ones, or if Anthropic’s earlier AECI growth was unusually slow versus the broader industry frontier. When METR plots frontier models beyond Anthropic alone, estimated acceleration falls toward roughly 1.4×—one reason the public should treat “twice as fast” as a lab-specific, method-dependent claim rather than a universal law of AI progress.
Saturation is another practical limit. METR notes that “Model 4,” an internal model finished around September–October 2026, scored above 50% on every AECI benchmark it was tested on. Measurement gets harder as the yardsticks fill up.
The redactions are part of the news
Anthropic had unilateral rights to redact non-public information. METR retained the right to say whether redactions happened and whether findings depended on them. Both did. Anthropic redacted “a majority of the evidence we found most informative,” describing the cuts as aimed at confidential information that could accelerate other AI developers.
What the public can see includes AECI trends, the no-compute-acceleration finding, and partial interview evidence. What it cannot see includes key intermediate capability measures, algorithmic-progress estimates, detailed attribution by pipeline stage, bottleneck analyses, forecasts for future models, and METR’s more granular acceleration estimates. METR is explicit about the consequence: the public must lean on METR’s judgment more than would be ideal. This is transparency about conclusions more than transparency about evidence.
Process details still matter for trust. The engagement used about 2.5 full-time-equivalent METR staff over eleven weeks. Anthropic did not pay METR; it provided devices and tokens. METR did not get employee-like, unmediated access or query access to unreleased models. A contracted compute-allocation verification—capabilities versus safety spend—was canceled mid-engagement before METR reached conclusions on that question. Anthropic said it wanted to keep the work focused on risk assessment and would revisit compute verification later.
Timing against “pace the frontier”
The evaluation’s calendar overlaps Anthropic’s public pacing message. METR began receiving confidential information on August 12. On September 12—the same day METR delivered a first full-report draft—CEO Dario Amodei published “We Must Pace the Frontier,” arguing that companies must slow the rate of capability improvement so safety work can keep up. His first proposed step was embedded third-party evaluators with employee-like access; he named METR as an example and said Anthropic was unilaterally committing.
This assessment is not that embedded program. METR distinguishes it from the employee-like evaluator access Anthropic discussed in a mid-October update. It is still the clearest public test so far of what “independent look at our pace” produces when a lab retains redaction rights and narrows scope. Anthropic co-founder Sam McCandlish has said the company plans its own analysis of internal progress speed. That response will be worth reading against METR’s uncertainty bands, not only against the 2× headline.
The political backdrop sharpens the stakes without changing the evidence. Midterm results and data-center backlash have already forced the industry into a more defensive posture on infrastructure and regulation. A credible finding that at least one frontier lab’s benchmark progress has sped up—while its CEO argues for pacing—will be used by people who want slower AI and by people who want faster deployment. The report itself supports a narrower claim: progress on a particular class of tasks appears to have accelerated at Anthropic; AI assistance is a leading explanation; compounding remains a live risk because METR did not see clear near-term bottlenecks; and the public cannot inspect most of the evidence that most moved the evaluators.
What to watch next
Three follow-throughs will decide whether this assessment was a one-off snapshot or the start of a verification regime.
First, whether Anthropic publishes its promised internal pace analysis with enough method detail to compare against AECI’s 1.9× and METR’s uncertainty range.
Second, whether future evaluator access moves from negotiated document sharing toward the employee-like embedding Amodei described—and whether redaction rules shrink when the question is public risk rather than competitor leakage.
Third, whether “signs of life” for autonomous contributions become measured shares of pretraining and reinforcement-learning work. Assistance can double a research team’s throughput. Autonomy that improves the training loop is the threshold where acceleration can feed on itself. METR says that threshold has not been crossed in the most important stages. It also says it did not see bottlenecks that would reliably prevent further acceleration over the next year.
The useful reading of the METR report is therefore neither panic nor complacency. Anthropic appears to be moving faster on well-specified capabilities than its recent past, AI tools are probably a major reason, and the evidence the public can audit is thinner than the importance of the claim. That combination—stronger pace, softer proof—is exactly what pacing rhetoric has to confront if it is going to be more than a brand position.