METR Says Anthropic Is Improving Its Models About 2x as Fast, With AI Driving Most of It

On November 5, 2026, METR published its estimate that Anthropic is now improving its models' capabilities at about twice the rate it did between March 2024 and October 2025, and that most of the speed-up, perhaps around 80 percent, comes from Anthropic using AI to do AI research. METR calls this a high-uncertainty judgment, and says Anthropic redacted a majority of the evidence behind it.

The claim comes from two documents: METR's independent assessment of AI R&D acceleration at Anthropic, and its review of the "Risks from automated R&D" section of Anthropic's August 2026 Risk Report. We read both. This post sorts what METR actually wrote from what the headlines compress.

TL;DR: the numbers and how much to trust them

What METR actually did

METR is a nonprofit that evaluates frontier models. From August 12 to October 29, 2026, about 2.5 full-time-equivalent METR staff ran what it calls a high-access evaluation at Anthropic. Anthropic shared internal benchmark scores for unreleased models, training compute metrics, headcount data, lines of code and pull requests, AI inference spend, surveys of self-reported uplift, and about 15 staff interviews.

METR says the access "surpassed our expectations," and that Anthropic shared more than was first scoped. It also says Anthropic could reject any request, and some were denied. METR worked on Anthropic devices, and says it took no payment from Anthropic, though Anthropic provided free tokens for AI-assisted analysis.

The engagement has a second output. Anthropic asked METR to review the AI R&D section of its August Risk Report. That review is published separately and is the source of the claim that the public report understates acceleration.

This is not the independent-evaluator program Anthropic announced on October 15. METR states the two are distinct.

The headline estimate: roughly 2x, with wide error bars

METR's executive summary says that, as of October 2026, Anthropic is likely improving model capabilities on well-specified tasks at around twice the rate it did between March 2024 and October 2025. It adds that it "would not be surprised" if the true figure were closer to 1.3x or 3x.

The anchor is a benchmark composite. AECI is Anthropic's internal version of the Epoch Capabilities Index. METR fit a trend with one breakpoint and found growth since October 2025 at 1.9x the earlier rate, with a 95 percent interval of 1.5 to 2.8x. In METR's words this implies Anthropic made roughly two years of historical progress over the last year.

Three cautions from METR itself:

• The data points lag the research, because a model only shows up after it is trained.
• AECI focuses on verifiable tasks, so it may overstate progress on messy real-world work.
• Anthropic's historical growth was lower than the industry frontier before late 2025, which could inflate the ratio.

METR also says the 1.9x figure is higher than the 1.7x in the Haiku 5.5 system card, mainly because METR included newer unreleased models. It tested robustness against choices about benchmarks, model dating and saturation, and got rates between 1.1x and 2.9x.

Where the "80 percent from AI" figure comes from

METR's section on attribution says the late-2025 acceleration was likely influenced by non-AI factors, and that its details are redacted. For the current acceleration it writes that it guesses a majority, "perhaps ≈80%," is due to AI.

That is a subjective estimate, not a measured share. METR says it rests on several lines of evidence and that Anthropic redacted its detailed attribution estimates, including the breakdown by pipeline stage. What readers can see is the supporting context:

• Interviews: researchers gave a wide range of estimates, centered around a 2x acceleration attributable to AI use.
• Slack polls: staff vote on speedup from the latest model, and self-reports reached 12x in October 2026. METR calls this very loose evidence from low-effort responses, and notes that an individual's speedup is not the same as an acceleration of overall progress.
• Autonomy: Anthropic staff told METR that AI mostly assists rather than contributes autonomously to the most important training work, but described "signs of life" that autonomous contributions could come soon.

The distinction matters. A 12x self-reported uplift does not mean 12x faster progress, since diminishing returns, compute and evaluation all bound the output.

The Risk Report dispute: "less than 2x" versus "around 2"

Anthropic's Responsible Scaling Policy v3.4 sets two thresholds for automated R&D in key domains. One is full substitution of its research scientists and engineers at competitive cost. The other, which METR labels the "2x Acceleration Threshold," covers dramatic acceleration of AI progress that likely relates to automation of AI R&D. Anthropic's August 2026 Risk Report concludes risk is low and that neither criterion was met as of July 15.

METR agrees on the first threshold and on the low-risk bottom line. It disagrees on how firmly the second can be called:

1. Leading indicators. The public report says acceleration was "less than a factor of 2." METR says it calculated 1.97 by combining figures from different parts of the private report, and that "around 2" would be the fairer description. METR's own range across definitions was 1.3 to 2.8.
2. Attribution to AI. The public report suggests a smaller AI share. The private report says it is plausible the AI-attributable counterfactual speedup exceeds 2x. METR says it broadly agrees with the private version.
3. Whether it was a one-time jump. The public report offers weak evidence for a one-time jump in AECI at Mythos Preview. METR says that, using the report's data, estimated acceleration was about 1.5x, and about 1.9x once Anthropic added more benchmarks after publication. It prefers the reading of continued acceleration.

METR's summary is that the public report "understates the level of acceleration," while both METR's researchers and Anthropic's staff lean toward the threshold not having been crossed. That is a disagreement about confidence, not a claim that Anthropic hid a breach.

What was redacted, and why it matters

This is the most unusual part. METR's assessment opens with a redaction notice: Anthropic had rights to redact any non-public information, METR had the right to say that it did, and "Anthropic redacted a majority of the evidence we found most informative." Anthropic described the aim as removing confidential information that could accelerate other AI developers.

METR published a table of what it saw versus what it could report. Benchmark capabilities of internal models were shared and reported. Intermediate measures, algorithmic progress estimates, and bottleneck and forecast details were shared, at least in part, but not reported. Evidence of AI's contribution and staff interviews were only partly reported.

METR is blunt about the consequence. Because the public cannot check the evidence, it says, "the public has to rely on METR's judgment much more than would be ideal." It also says the report does not achieve transparency about evidence.

Two further process facts are on the record. Anthropic canceled a planned component that would have verified how its compute is split between capabilities and safety, before METR reached conclusions, and said it would revisit it later. And METR's draft sat through a roughly five-day extension while Anthropic tightened redaction guidance in late October.

Bottlenecks and the six-month warning

METR says it did not see evidence of significant bottlenecks in algorithms, training compute, experiment compute or data that would bite in the next 12 months. It adds that its view of bottlenecks changed several times as more detail arrived, so it holds that view loosely, and that Anthropic redacted its analysis of specific bottlenecks.

In the Risk Report review it goes further: it cannot rule out a substantial further acceleration over the next six months, which "could rapidly lead to 'high' risks." That is METR's forecast stance, not Anthropic's.

Why this is more than an Anthropic story

AI labs have long talked about models helping build the next models. Until now the evidence was mostly company self-report. This is the first time an outside group with internal data access has put a number on it publicly, and the number sits right at the 2x line that Anthropic's policy treats as an early warning.

Three readings are worth holding at once:

• It is consistent with what Anthropic has said. Its S-1 amendment signaled its next models may imminently meet the accelerated-progress threshold, which we covered as a claimed-versus-verified post.
• It fits the wider field. OpenAI published its own account of research acceleration in September, which we broke down in OpenAI's Research Acceleration post. Geoffrey Hinton's argument for an intelligence explosion is summarized in our Hinton paper explainer.
• It lands during the IPO process. Anthropic's filing and its risk disclosures are under scrutiny, as in our read of the S-1 and Senator Warren's questions to Amodei. A third-party estimate near the threshold is material context for investors and regulators, though METR itself makes no financial claim.

What to take from the numbers: claimed versus verified

Nothing here shows AI doing research without humans. Both organizations say it is assistance for now. The conclusions are also about capabilities on well-specified tasks, the kind that are easy to grade automatically. METR lists the opposite risk too: the acceleration may be overfit to the benchmark mix, so real-world impact could be smaller.

What builders and readers can do

The practical effect for most people building with Claude is indirect, but it is real.

• Expect faster model cadence. If the estimate holds, new Anthropic models will arrive sooner than the 2024-25 rhythm suggests. Plan migrations and evaluation harnesses for shorter cycles. Our look at the Fable 5.1 and Mythos 5.1 launch shows the current baseline.
• Read the thresholds, not only the headline. Check how a lab defines "crossed." METR's critique is partly about phrasing: "less than 2" versus "around 2."
• Treat self-reported uplift with care. Poll numbers like 12x are not measures of output. Ask for held-out evaluations.
• Watch what outside reviewers can actually see. Whether Anthropic releases more of the redacted material, and whether the compute-allocation check happens, are the two things to follow.
• Secure your own agents. As labs use agents on their own research, the same agent-risk questions apply to your stack. If you run coding agents, a tool such as AgentBeam, from the explainx.ai team, stops AI agents before they take dangerous actions. For a related case, see Anthropic's self-replicating AI test on IoT devices.

Agents that act on your behalf need rules. See how explainx helps secure your agents →

Limits of this analysis

We read METR's two posts and the structure of Anthropic's report, not METR's roughly 100-page unredacted full report, which only Anthropic has. The numbers above are METR's, and we have not independently reproduced them. METR's own limitations section is worth reading in full: it says fuller access would have sped up proper interpretation, and that it reversed itself several times on critical questions as context arrived.

A green lens inspecting a labeled card, illustrating how an outside group audits an AI lab's report

Related reading

• Anthropic's S-1 amendment and the accelerated-progress threshold
• OpenAI's Research Acceleration post: 3.1 agent-workdays per human
• Hinton's intelligence explosion paper
• Anthropic's S-1: AI bubble and the future of work
• Warren's letter to Amodei on the S-1
• Claude Fable 5.1 and Mythos 5.1 launch
• Anthropic Q3 revenue of 16.8B: claimed vs verified

Figures and quotes are accurate as of November 5, 2026, based on METR's published posts and Anthropic's redacted August 2026 Risk Report.