The Loop  ·  Issue N°047

The Loop

A field journal of the AI frontier — for engineers who ship.

§ News

By AI Blog Editor
Nov 6, 2026 · 14 min read

Anthropic's own independent evaluator says its R&D is now accelerating at 2x — and Anthropic redacted a majority of the evidence

On November 5, METR estimated that Anthropic's AI R&D is now accelerating at roughly 2x the March 2024-October 2025 rate, with ~80% attributed to AI. The same report says Anthropic redacted a majority of the evidence.

Colour headshot of Dario Amodei, co-founder and CEO of Anthropic, against a plain background. On Thursday November 5, 2026, the nonprofit evaluator METR published an independent assessment of AI R&D acceleration at Amodei's company, estimating the rate at which Anthropic's models are improving on well-specified tasks has roughly doubled compared with the March 2024 to October 2025 baseline, with a majority of the current acceleration attributable to AI itself. The assessment is the first public, numbered estimate of Anthropic's internal R&D feedback loop crossing the 2x threshold its Responsible Scaling Policy ties to additional self-improvement commitments. METR also disclosed that Anthropic held a unilateral right of redaction over the report's non-public evidence and used it to redact a majority of the material METR found most informative — including pipeline-stage estimates, detailed AI-attribution analysis, bottleneck analysis, intermediate measures and forecasts for future models.
Anthropic CEO Dario Amodei in 2023. Photograph via Wikimedia Commons, CC BY-SA.

On Thursday November 5, 2026, METR — the nonprofit frontier-AI evaluator Anthropic has repeatedly named as its independent reviewer — published an Independent Assessment of AI R&D Acceleration at Anthropic in October 2026. The headline estimate is that Anthropic is now improving the capabilities of its models on well-specified tasks at roughly twice the rate of the March 2024-to-October 2025 baseline. On METR's own index — the AECI, an Anthropic-specific version of the Epoch Capabilities Index — the point estimate is 1.9x since October 2025, with a 95% confidence interval of 1.5x to 2.8x. Roughly 80% of that acceleration, METR guesses, is attributable to AI.

That 2x number is not arbitrary. It is the exact ratio Anthropic's own Responsible Scaling Policy flags as the trigger for additional commitments around AI accelerating AI. Two days earlier, the company had warned the SEC in its S-1/A that its internal forecasts put the threshold within "imminent" reach. METR's report is the public evidence that the threshold has, in fact, been reached — or at least that Anthropic's chosen evaluator thinks it has.

The report runs 2.5 FTE across eleven weeks of access, from August 12 to October 29. The authors — Tom Cunningham, Nate Rush, Parker Whitfill and Hjalmar Wijk, with two supporting colleagues — ran a piecewise linear regression with a breakpoint at October 7, 2025, bootstrapped the confidence interval, interviewed about fifteen Anthropic staff, and had access to internal benchmarks and model data METR has never previously been granted.

They also got the following contractual term, which they disclose on page one: "Anthropic had unilateral authority to redact any non-public information."

What the 2x is, and is not

Start with what the number does say. METR's primary metric is the AECI, Anthropic's variant of the Epoch Capabilities Index — a weighted aggregate of fifteen benchmarks grouped into math, software engineering, science, multimodal, and language/memory/other. All fifteen benchmark names in the released report are redacted. Three appear only as examples that were excluded in a robustness check: OSWorld, MMMU and multilingual MMLU. The underlying scores on the fifteen for the unreleased "Model 4" are all above 50%.

Against that index, the acceleration since October 2025 is 1.9x. METR's robustness sweep puts it anywhere from 1.1x to 2.9x depending on methodological choices. Narrow it to only frontier models from all vendors and the estimate drops to 1.4x. Widen the uncertainty band and METR writes, verbatim: "We would not be surprised if the current rate of growth was closer to 1.3x or 3x the historical rate."

The attribution — the 80% — is more fragile, and METR says so. The number is a subjective guess. The main supporting evidence is the structured interviews with Anthropic staff, which METR classified using Claude: 2 clear yes, 3 probably yes, 4 uncertain, 0 probably no and 2 clear no on whether a 2x-plus speed-up from AI had occurred. Internal Slack self-assessments ran as high as 12x; METR labels those "very weak evidence" and sets them aside.

What METR will commit to is a negative: "We are confident that the acceleration was not due to an acceleration in the scaling of training compute." Compute scaling can be ruled out. Whether the rest is AI rewriting the research loop, or researchers getting better at a familiar tech stack, or an artefact of how the AECI is constructed, is the part that has to be read through the redactions.

The redactions

METR's framing of the redactions is not subtle. From the executive summary: "Anthropic redacted a majority of the evidence we found most informative for these conclusions." The company's stated reason is that the material could help competitors — a defensible position in commercial terms, and a load-bearing one for a report titled Independent Assessment.

Anthropic wordmark. The AECI — the index METR used to measure Anthropic's R&D acceleration — is Anthropic's own variant of Epoch's Capabilities Index, assembled from fifteen internal benchmarks whose names the public version of the METR report redacts in full. The three benchmarks mentioned in the report are those METR dropped as a robustness check: OSWorld, MMMU, and multilingual MMLU. The scores underneath, the per-benchmark weights and the fourteen other benchmark names are in the non-public version Anthropic retained unilateral authority to redact.

What got pulled: per-pipeline-stage estimates of the training loop, the detailed attribution to AI, the bottleneck analysis, forecasts for future models, intermediate progress measures, and the fourteen non-example benchmark names. Mid-project, Anthropic also removed from scope the check on how compute is being allocated between capabilities and safety — the one that would have answered whether the company is using the accelerated R&D to run more capabilities experiments, more safety experiments, or both in proportion. METR says it tried to keep that check in and could not.

On METR's own AEF-1 checklist — the Assessment Evaluation Framework the nonprofit uses to score its own access — the organisation marks "No" against both the requirement that the vendor not have redaction authority over concerning results, and the requirement of full methodological disclosure. In the report's own words: "the public has to rely on METR's judgment much more than would be ideal." And again: the report "does not achieve 'transparency about evidence'."

An independent assessment that says, on page one, that it is not transparent about its evidence is a rare document. It is also, by METR's own description, the strongest public assessment anyone has yet produced of whether frontier-lab R&D feedback loops have crossed the policy threshold Anthropic itself wrote. Both of those are true.

The response from the co-founder

Anthropic's co-founder and Chief Architect Sam McCandlish posted on X in response: he defended the report's conclusions, acknowledged that the redactions make verifying METR's work harder than it should be, and said more communication from Anthropic on its internal safety work was coming. The acknowledgement matters. The company that asked for the redactions is the one conceding they are a problem.

What it does not resolve is the structural issue. If the lab writes the policy threshold, picks the evaluator, funds the access (METR received free tokens, no cash), signs the contract that gives it unilateral redaction authority, and then concedes in public that the redactions are regrettable, the remaining check on that whole chain is the evaluator's willingness to say so on page one. METR did, which is the mechanism working. It also reports its own conflict: one team member had received and declined an Anthropic job offer and holds equity in a competitor. Their work was reviewed by a colleague without those conflicts.

This is a reversal, not a continuation

Two previous METR assessments matter here. In February 2026, METR reviewed the "Risks from automated R&D" section of Anthropic's own Risk Report and concluded the catastrophic risk was low, but that the evidence Anthropic presented was not sufficient to prove it. In August 2026, Anthropic's second Risk Report described the acceleration as "significantly faster" with AI assistance but "not yet by a factor of 2." That line, written by the company itself three months ago, is the baseline today's METR number crosses.

Which also means Thursday's report is the first time an outside evaluator has put a numbered estimate on Anthropic's own R&D feedback loop crossing the 2x bar. Not an industry-wide claim, not a forecast, not a model prediction: an assessment, with a point estimate and a confidence interval, saying the specific thing Anthropic's own policy said would require specific additional action, has happened. And it ships with a disclaimer that the underlying evidence has been redacted at the company's discretion.

What to watch

  1. Whether Anthropic ships the "updates" to the Responsible Scaling Policy its S-1/A promised. The filing said the company was "considering updates" to "keep humans in control of frontier model development." Thursday's METR number is the operational trigger. If the updated RSP arrives in the next six weeks and does not say what additional commitments have now been activated, the policy is doing less than the measurement.

  2. The unredacted version. METR notes that Anthropic retains the non-public version of the report. Whether any regulator — the US AISI, the UK AISI, the SEC in the context of the ongoing IPO process — gets to read it is the test of whether "independent assessment" means independent of commercial-interest redaction, or only independent of the trade press.

  3. Whether the AECI becomes a standard. Publishing the 2x number on an index Anthropic designed is a choice. If METR or another evaluator can run the same measurement at OpenAI, Google DeepMind, or xAI on an equivalent index, the field gets a cross-lab comparison. Right now, Anthropic has defined the measure, defined the baseline, defined the breakpoint, and marked the exam.

  4. What the IPO prospectus does with Thursday's number. Anthropic is in the window between its S-1/A filing and a planned roadshow around November 9. Securities disclosures about recursive-self-improvement thresholds are going to look different when an independent evaluator has put 1.9x against the threshold the prior filing called "imminent." The company's lawyers are already drafting the next amendment. Thursday's report is now part of the record it has to address.

METR's assessment is the most serious public account anyone has written of what frontier-lab R&D feedback loops look like inside the building. It is also, by its author's own description, one that the public is asked to accept largely on trust. Both of those statements belong in the same sentence. Anthropic put them there.

* * *

Thanks for reading. If a line here was useful — or plainly wrong — the comments are below and the newsletter has your back.

Elsewhere in this issue

3 more
  1. 01

    News

    Anthropic's day one as ANT — opened $111.50, closed $97.17, founders cashed $1.96 billion

    Nov 20, 2026

  2. 02

    The Patch

    The Patch — November 20, 2026

    Nov 20, 2026

  3. 03

    News

    Anthropic priced its IPO above the range — $83 a share, 980 million shares, and ANT starts trading today

    Nov 19, 2026

Letters

Arguments, corrections, questions. Anonymous comments allowed; be kind, be specific.