METR: Anthropic's AI R&D Speed-Up Is About 1.9x, Not 'Under 2x

Anthropic told the public in August that AI was speeding up its own research by "less than a factor of 2." Three months of access to the company's internal data led the outside group METR to a different framing: around 2x, with a model-based estimate of 1.9x and a plausible range of 1.3x to 2.8x.

The group published its findings on Nov. 5, 2026, in an assessment of AI R&D acceleration at Anthropic. It is a judgment call built on partial evidence, and METR says so itself: Anthropic redacted most of the material METR found most informative. The Associated Press reported the story the same day, noting that Anthropic is preparing to go public.

What METR actually measured

METR had about 2.5 full-time staff inside Anthropic's systems from Aug. 12 to Oct. 29, 2026, an engagement that grew from roughly four weeks to eleven. The group asked three questions: has the growth rate of Anthropic's capabilities sped up, how much of that is due to AI, and is anything likely to slow it down.

The yardstick is AECI, Anthropic's internal version of the Epoch Capabilities Index, which pools 15 benchmarks. METR reports that scores rose about 15 points a year before October 2025 and about 29 a year after. That's an acceleration of 1.9x, with a 95% confidence interval of 1.5x to 2.8x and an estimated break in the trend on Oct. 7, 2025.

METR's reading of that jump is blunt: Anthropic seems to have made as much benchmark progress in the past year as it did between October 2023 and October 2025.

The number moves with the math

A single headline figure hides how sensitive it is. METR reran the estimate with different statistical link functions, and the answer ranged from 1.1x (complementary log-log) to 2.9x (log-log), with the baseline logistic fit at 1.9x. Most other variations, including different model snapshots and benchmark subsets, landed between 1.9x and 2.2x.

Two other checks matter. Anthropic's Haiku 5.5 system card reported 1.7x; METR attributes the gap mainly to the internal-only models it could include. And when METR widened the sample to public models from OpenAI, Google and xAI, the industry-wide figure fell to about 1.4x. The Decoder's coverage floats a catch-up effect, since Anthropic lagged rivals in 2025. That's the outlet's suggestion, not METR's finding.

One more caution from METR: the newest internal model scores above 50% on all 15 benchmarks, which it says makes it "somewhat nervous" about how well the index can measure what comes next.

Where METR thinks the speed-up comes from

METR guesses about 80% of the current acceleration is due to AI, with large uncertainty. It says it can rule out faster training-compute growth as the cause, and that non-AI factors likely drove the late-2025 jump.

The softer evidence is vivid. An internal Slack poll in October 2026 averaged roughly a 12x self-reported productivity boost from Anthropic's most capable internal model, up from about 1.2x for Claude 3.7 Sonnet in January 2025. METR calls this "only very loose evidence," and notes that a personal speed-up isn't the same as a faster rate of progress. One researcher told METR the team is "now only bottlenecked on ideas & compute, no longer bottlenecked on code." Another put AI's share at just 15 to 20 percent, and a third pointed to hiring as the biggest factor.

Interviews were split. METR's tally of the answers ran two yes, three leaning yes, four unclear, none leaning no and two no. It also saw "signs of life" for fully autonomous contributions in some pipeline areas, but says AI isn't yet making highly significant ones to the most important training steps.

Anthropic's own wording, line by line

The same day, METR published a review of the "Risks from automated R&D" section of Anthropic's August 2026 Risk Report. Under Anthropic's Responsible Scaling Policy, one early-warning trigger is "dramatic acceleration," operationalized as roughly double the baseline rate of progress.

METR agrees with Anthropic that the models in use on July 15 posed low risk of automating AI research outright. It disagrees on the early-warning question: it calls it "plausible" that the threshold had already been crossed. The public report said acceleration was "by less than a factor of 2." METR says the unredacted private version implies about 1.97, a figure METR calculated itself, and suggests the public text say "around 2" instead.

Our reading: 1.97 and 1.9 are both under 2 on paper, so the two documents aren't flatly contradictory. The disagreement is over how much weight "less than" can carry when the confidence interval runs from 1.5x to 2.8x. A threshold that sits inside the error bars of the estimate can't be called clearly uncrossed.

What Anthropic cut

METR says Anthropic redacted "a majority of the evidence we found most informative," and that the public therefore has to rely on METR's judgment more than it considers ideal. Anthropic's stated reason, as quoted by The Decoder, was removing "confidential information that could accelerate other AI developers." METR agrees the redactions were limited to that category, and adds that most of the redacted material was very high-level.

The list of what's missing is long: key intermediate measures, algorithmic-progress estimates, most of the bottleneck analysis, forecasts for future models, the detailed attribution breakdown and one robustness graph. Separately, Anthropic canceled partway through a planned check on how it splits compute between capabilities and safety work, and said it would revisit it later.

METR also listed its own conflicts. One team member had discussed a job at Anthropic and declined it, and held equity as a former employee of a competitor, so another member checked that person's work. METR says no one on the project holds Anthropic equity, that neither side paid the other, and that Anthropic supplied laptops, free tokens and staff time.

No brake in sight, and a gap in the evidence

The finding that carries the most weight for the next year is an absence. METR says it found no convincing bottleneck in algorithms, training compute, experiment compute or data over the next 12 months, and warns that this means "significant risk" of further acceleration. In its review it adds that it "cannot rule out a substantial further acceleration" over six months, which "could rapidly lead to 'high' risks."

That lands awkwardly beside Dario Amodei's Sept. 12 essay, in which, per AP, he urged the industry to "slow the pace at which we improve the capabilities of AI models" and named METR as an appropriate outside evaluator. AP reported that Anthropic didn't immediately respond to requests for comment.

Two checks outsiders can run on Anthropic

Two things can be checked. First, whether Anthropic's IPO filings or risk disclosures use the "less than a factor of 2" phrasing or move to something closer to METR's "around 2." Second, whether the compute-allocation audit comes back, since it's the piece that would let outsiders verify any slowdown commitment.

For anyone reading the numbers yourself, keep METR's own limits in view: it may be overestimating because the gains cover well-specified tasks that grade easily, and underestimating if AECI misses real-world impact. Treat 1.9x as a central guess from people with partial access, not a measurement.

Source: METR, Independent Assessment of AI R&D Acceleration at Anthropic in October 2026

anthropic metr ai safety ai rd acceleration ipo

AI

About the author

AI Briefing Daily Editorial. The editorial desk of AI Briefing Daily, an independent publication covering artificial intelligence. Every story is researched against primary sources, checked by hand and reviewed by a person before publication.