METR finds Anthropic's AI research pace is about 2x faster, with most of the evidence redacted

METR, the AI evaluation nonprofit, published an outside assessment of Anthropic on November 5, 2026, and its central finding is the one the superintelligence debate has been waiting for: a measurable speedup in how fast a frontier lab improves its models, and a guess that AI itself is doing most of the work. METR estimates Anthropic is improving capabilities on well-specified tasks at about twice its March 2024 to October 2025 rate. The catch is that Anthropic decided what the public could see.

What METR actually did

The evaluation ran from August 12 to October 29, 2026, using about 2.5 full-time-equivalent staff over eleven weeks, according to METR. The team looked at internal benchmark scores, Anthropic's own version of the Epoch Capabilities Index (AECI), algorithmic efficiency estimates, headcount, training compute, lines of code, inference spend, surveys and about 15 staff interviews. The final report was submitted November 3.

METR asked three questions. Did Anthropic's capability growth accelerate? How much of any acceleration is due to AI? And could bottlenecks limit growth over the next year?

METR did not get query access to unreleased models, and a planned check on compute allocation was canceled partway through the engagement.

The numbers

Since October 2025, METR estimates AECI growth at 1.9x the earlier pace, with a 95% interval of 1.5x to 2.8x. In index terms, the trend moved from about 15 points per year to about 29. Across alternative methods the estimates ranged from 1.1x to 2.9x, and METR says it "would not be surprised if the true acceleration was closer to 1.3x or 3x the historical rate."

For comparison, METR notes that the Haiku 5.5 system card reported 1.7x, and that the frontier trend across all developers shows about 1.4x. METR's own phrasing is that Anthropic "has made roughly two years of historical progress over the last year."

Why METR thinks AI is the cause

METR rules out the boring explanation first. It is confident the speedup did not come from faster scaling of training compute. Its subjective guess is that perhaps 80% of the current acceleration is due to AI, with large uncertainty.

The worker evidence is softer. Informal Slack polls put average self-reported uplift near 12x for the most capable internal model in October 2026. Interviews were far more cautious. Of 12 interviewees, 2 were coded "yes" on whether AI supports more than 2x acceleration, 3 "leans yes," 4 "borderline or unclear," none "leans no" and 2 "no." One researcher told METR they "could definitely endorse a greater than 2X speedup." Another said they do not believe in self-reported productivity increases.

METR also flags a measurement problem: the model it calls Model 4 scores above 50% on all 15 AECI benchmarks it used, which means the yardstick may be running out of room.

The redaction problem

The most important caveat sits at the end. Anthropic had unilateral redaction rights and, METR says, redacted most of the evidence METR found most informative. Anthropic described the redactions as aimed at removing confidential information that could accelerate other AI developers. METR's conclusion is that the public "has to rely on METR's judgment much more than would be ideal."

METR also lists reasons its estimate could be too high: non-AI factors could explain much of the speedup, gains may be limited to automatically graded tasks, and the earlier baseline may have been unusually slow. It lists reasons it could be too low as well, including benchmark gains that understate real-world impact. METR considers deliberate misleading by Anthropic extremely unlikely, but notes most analysis relied on Claude and that more access would strengthen robustness. It also disclosed that one team member had been in discussions about a role at Anthropic, declined it, and had former equity in a competitor.

How to read the number

Two cautions help. First, "twice as fast" refers to an index of capability on tasks with clear answers, not to a lab's ability to solve open research problems. Second, the baseline matters: METR compares against Anthropic's own March 2024 to October 2025 pace, and notes that period may have been unusually slow. A faster comparison period would shrink the headline.

The comparison with other labs is also incomplete. METR cites the all-developer frontier trend at about 1.4x, which suggests Anthropic is ahead of the field but not by an order of magnitude. Whether the gap persists depends on whether competitors adopt similar AI-assisted workflows, which is a question no outside evaluator has yet been able to measure at them.

No bottlenecks, with low confidence

On whether something will slow the curve, METR says it "did not see evidence for significant bottlenecks in algorithmic progress, training compute" over the next 12 months, while stressing low confidence and noting its understanding changed several times during the work.

What it means for the race

This is the first outside estimate of AI-assisted research acceleration at a named frontier lab, and it lands in the middle of the argument covered in our explainer on recursive self-improvement. A 2x pace is not an intelligence explosion. It is also not nothing: it is what the lab's own staff and an outside evaluator both see in the data, on tasks that can be graded automatically.

It also extends METR's earlier role at the lab, which we covered in our review of METR's look at Anthropic's risk report. The pattern is consistent: access improves each round, redactions remain.

Our position: believe the direction, discount the precision. The useful test is whether other labs allow similar assessments with fewer redactions, and whether a second METR pass with access to unreleased models lands near 2x or nearer 3x. Until then, the honest summary is that a major lab is probably getting faster, mostly because of its own AI, and the public cannot see the strongest evidence.