They redacted the evidence
METR spent eleven weeks inside Anthropic measuring AI R&D acceleration at about 2x, then watched the lab redact most of the evidence the evaluators found most informative. The public is asked to trust blank pages; Amodei's pacing essay landed mid-engagement, and McCandlish defended the cuts on X.
There is a sentence in a public report that asks you to trust a blank page. On Thursday, METR published an independent assessment of AI R&D acceleration at Anthropic and put the cost of that trust in the open. Four researchers — Tom Cunningham, Nate Rush, Parker Whitfill, and Hjalmar Wijk — had spent eleven weeks, from August 12 to October 29, inside the lab with high access, on Anthropic devices, with free tokens for analysis and no payment from the company. They delivered a roughly one-hundred-page Full Report to Anthropic on October 22. Anthropic sent redaction guidance on October 25, tightened it again between October 27 and October 29, and METR submitted the final public version on November 3. What the public got on November 5 was the high-level risk conclusions and a table of what had been cut. In METR's own words, posted the same day: "We submitted a ~100-page full report on our assessment to Anthropic. For the public version, Anthropic redacted a majority of the evidence we found most informative. As a consequence, the public must rely on our judgment more than we consider ideal."
I keep coming back to that last clause, because it is the part that rearranges the furniture. The headline findings are striking enough on their own. As of October 2026, METR's subjective view — given with high uncertainty — is that Anthropic is likely improving capabilities on well-specified tasks at around twice the rate it managed between March 2024 and October 2025. The team would not be surprised if the true figure sat closer to 1.3x or 3x. A majority of the current acceleration, they guess, is attributable to extensive AI use in R&D; their guess lands near eighty percent of the October 2026 acceleration, with substantial uncertainty still attached. They did not see clear bottlenecks in algorithmic progress, training compute, experiment compute, or data for the next twelve months, and they flag a significant risk that further capabilities could produce still faster acceleration. Those are the sentences the public is allowed to keep. The evidence METR says was most informative for reaching them is mostly gone.
Eleven weeks inside
The engagement was not a light look at a system card. Anthropic offered what METR describes as unprecedented access to internal data on the AI R&D process: metrics on pretraining and reinforcement learning, training compute by domain, benchmark scores for internal models, AI speedup analyses, forecasts for a future model's metrics, and interviews with relevant staff. The team worked as about 2.5 full-time equivalents. They could request anything they believed relevant; Anthropic could approve or reject unilaterally, and in practice the two sides negotiated toward data that was informative, not extremely sensitive, and easy to share. Access surpassed METR's expectations. It was still not employee-like access. Assessors did not get unmediated runs through documents, Slack, or non-public models, and they never requested sample access to unreleased systems. The original contract had also included a verification of how much compute went to capabilities versus safety. Anthropic canceled that mid-engagement, before METR reached substantive conclusions on allocation, saying it wanted to keep the work focused on risk assessment for now and would revisit compute verification later.
What they could see of the pipeline was still a lot. METR reports that Anthropic shared extensive evidence across pretraining and RL, including estimates of algorithmic progress over time, and that this nonpublic evidence was highly informative about acceleration. They interviewed roughly fifteen researchers. They saw headcount, lines of code, pull requests, agent inference spend, internal surveys of self-reported uplift, and documents on autonomous research contributions. Two Anthropic staff whom METR describes as the highest-context researchers on measuring R&D acceleration met them frequently, collected data, ran analysis, and gave feedback. The Full Report tried to answer, in detail, the high-level questions METR had recently argued the public needs answered about AI R&D acceleration and loss-of-control risk. The public report is what survived Anthropic's unilateral redaction rights over non-public information. METR kept the right to say that redaction happened, and whether findings depended on what was cut. That is the bargain the industry currently calls independent evaluation.
Roughly two years in one
The cleanest unredacted number sits on the benchmark trend. Since October 2025, Anthropic's AECI scores — Anthropic's variant of the Epoch Capabilities Index — have been growing at an estimated 1.9 times the pre-October-2025 rate, with a ninety-five percent confidence interval of 1.5x to 2.8x. In plain language, that is roughly two years of historical progress packed into the last year. The learned breakpoint in the main specification lands around October 7, 2025. METR is confident the acceleration is not coming from faster scaling of training compute. Further details of that evidence were redacted.
Robustness checks matter here, and METR published them. Varying the link function that maps latent strength to benchmark scores swings the acceleration estimate from 1.1x to 2.9x. Snapshot choices, benchmark subsets, and regularisation leave the central result near 2x. The estimate sits above the 1.7x that appeared in the Haiku 5.5 system card mainly because METR included newer internal-only models: Model 2, finished in July 2026; Model 3, finished in August and September; Model 4, finished in September and October. Model 4 scored above fifty percent on all fifteen AECI benchmarks it was tested on. That saturation detail is the kind of quiet warning that usually lives in an appendix. If the hard end of the ruler is running out, measuring the next doubling gets harder even as the thing you are measuring may be speeding up.
Interviews and surveys lined up with a doubling story without settling the attribution cleanly. METR's overall impression of researcher estimates centered around 2x acceleration from AI use in their areas, with a wide range underneath. Claude-assisted coding of interview notes produced a mild lean toward yes on greater-than-2x acceleration in the interviewee's own work: two yes, three leans yes, four borderline or unclear, two no. Anthropic also shared Slack-poll self-reports in which staff reacted with estimates of speedup from the latest model. Those informal averages reached about twelve times by October 2026. METR flags that as very loose evidence — low-effort, unrepresentative, and historically prone to overstatement — and stresses that individual uplift is not the same thing as org-wide acceleration. Sustaining faster capability growth against diminishing returns usually needs a sustained change in the growth of inputs, not a single impressive multiple next to someone's name. AI is not yet making highly significant, fully autonomous contributions to the most important training stages. Researchers described "signs of life" that those contributions could arrive soon, with maturity varying by area. The more detailed estimates of how AI and non-AI factors split across pretraining, RL, and the rest of the pipeline were redacted.
The majority that cannot be shown
Attribution is where the blank pages do the heaviest work. METR's subjective view is that late-2025 acceleration was likely also shaped by non-AI factors — hiring, data, algorithmic discoveries, resolved bottlenecks — and that many of those details are among the redactions. For October 2026, their guess is that a majority of the current acceleration, perhaps around eighty percent, is due to AI, with substantial uncertainty and with AI's relative contribution growing over time. They investigated intermediate capability measures, algorithmic progress estimates, key R&D inputs, and bottleneck details. The public table is blunt about the outcome. Internal benchmark capabilities: shared with METR, described in public. Key intermediate measures: shared, not described. Algorithmic progress estimates: shared, not described. Evidence of AI's contribution: shared, only partly described. Measurements of key R&D inputs: mostly shared, not described. Details of bottlenecks and forecasts: partly shared, not described. Interviews: shared, only partly described. Anthropic described the cuts as aimed at removing confidential information that could accelerate other AI developers. METR agrees the redactions stayed inside that category, notes that much of what was cut was still high-level, and declines to referee the trade-off between a company's commercial interest and the public's interest in understanding risk. The evaluators' job, as they frame it, is to describe what was removed and how that shapes what you can conclude.
That is an honest institutional posture and a difficult civic one. Much of the redacted evidence, METR says, had multiple plausible interpretations. Both METR and Anthropic researchers carry significant uncertainty about how to synthesise it. The public cannot easily judge the evidence, reconcile it with outside sources, or reach its own conclusions. The report does not achieve what METR calls transparency about evidence. You get transparency about the existence of evidence, and then you are asked to trust the judgment of the people who saw it. Sam McCandlish, one of Anthropic's co-founders and a lead on its safety work, answered that bargain on X the same day. Many of METR's conclusions match his own. "AI is speeding up our research, but it's hard to say exactly how much," he wrote. "While they conclude AI can't yet make the most important contributions to model training on its own, they think the pace of development could, by default, accelerate further." On the redactions themselves he was direct: "We redacted a lot, and I accept that this makes it harder for others to check METR's work. But we determined that sharing exactly what is driving the speedup isn't the right thing to do, both because it's commercially sensitive and because it risks handing out a roadmap for others to go faster. That would make the risks METR is worried about worse. That's a tradeoff. I think we made the right call, though we're continuing to assess the pros and cons here." In a reply to Jerusalem Demsas he put the short version even more cleanly: "The short answer is that the unredacted information would allow others to speed up their research, and we thought the safety implications were worse than the benefits of transparency." The longer answer points to embedded third-party evaluators — more access and time for a reviewer the public is asked to trust, without making the underlying information public. In the coming weeks, he said, Anthropic plans to publish its own analysis of the speed of progress inside the lab, including where it stands against its Responsible Scaling Policy, plus new steps on safety and security as models keep improving quickly.
Pace, said while pacing
The calendar is doing some of the arguing on its own. Mid-September, while METR was still iterating on drafts inside Anthropic, Dario Amodei published "We Must Pace the Frontier," the essay that asked the industry to take the rate of advance seriously as a governance problem. The Verge noted the overlap without needing to invent irony: METR had been in communication from mid-August; the final public report was submitted November 3; the pacing essay landed in the middle. I do not need Amodei to have intended a contradiction for the juxtaposition to matter. A lab can argue, sincerely, that frontier progress should be paced and that the specific dials of its own acceleration are too useful to rivals — and too load-bearing for safety — to put in a PDF. A co-founder can accept that heavy redaction makes outside checking harder and still call the trade-off correct. An evaluator can say the public must rely on its judgment more than is ideal and still publish the judgment. All of those positions can be held at once. What they jointly produce is a particular kind of public knowledge: a risk claim with a confidence interval, a guess at AI's share, a table of missing rows, and a request that you treat the people who saw the missing rows as a Schelling point for trust.
That is the shape Thomas Wolf was pushing on when he argued that the report looks inward, and that labs themselves have become the points around which money and attention organise. Rohan Paul and others on X framed the moment as recursive self-improvement moving from theory into something you can at least attempt to measure. Accounts watching Model 4 and the cluster of unreleased systems treated McCandlish's defence of the redactions as the tell that the interesting graph was the one you cannot see. I am not going to invent quotes for any of that colour. The primary documents already say enough. RSI is no longer only a speculative story about a loop that closes someday. It is a measurement problem with a breakpoint in early October 2025, a confidence interval that still spans from modest to alarming, an eighty-percent guess that cannot show its working, and a lab that will publish its own speed analysis on a timetable of its choosing.
What trust looks like when the evidence is cut
I do not read METR's post as a failed audit. The access was real. The independence claim is serious: no payment, Anthropic devices, free tokens, unilateral rights to disclose that redaction happened. The AECI work is careful about link functions, saturation, and the difference between well-specified tasks and messier ones. The limitations section is unusually willing to list the ways the team could be overestimating or underestimating acceleration, including the possibility that Anthropic's earlier AECI growth was unusually low, that benchmarks overstate real-world impact, or that non-AI factors still explain most of the slope change. The honest difficulty is downstream of all that care. If the evidence that most moved the evaluators cannot travel with the conclusion, then the public sphere is being asked to do something closer to credentialed belief than to inspection. McCandlish's reply names that shift without flinching: trust in the evaluator rather than trust in the lab alone. That is better than trust in the lab alone. It is still trust.
The same Thursday offered a smaller rhyme from the other side of the frontier. OpenAI published a National Community Compact and a Compute Campus Policy Blueprint for future campuses it owns and operates — early resident input, local compact, paying full freight on electricity and infrastructure, water and noise reporting, local hiring, no NDAs for public officials except narrow security and trade-secret carve-outs — after a midterm cycle in which voters across dozens of jurisdictions backed data-centre restrictions or bans. That document is about power plants and school meals, not about AECI breakpoints, and I am not going to pretend it is the same story. It is a lab answering a trust deficit by publishing commitments the public can check without a cleared evaluator in the room. Put it next to Anthropic's redaction table and you get two theories of how a frontier company manages the gap between what it knows and what it will show. One theory hands the sensitive graph to a third party and asks you to trust their summary. The other publishes the operating rules for the campus and asks you to watch whether the meters match. Both are responses to the same pressure. Only one of them leaves the most informative evidence on the cutting-room floor by design.
Here is the part I think matters, and it is not that Anthropic is roughly twice as fast on well-specified tasks as it was. Plenty of people inside the labs already believed some version of that, and McCandlish says many of METR's conclusions match his own. The part that matters is the institutional shape of the disclosure. A nonprofit spent eleven weeks on the inside, wrote a hundred pages, kept the high-level risk sentences, and then watched the lab remove the majority of what those sentences rested on, because publishing the drivers of the speedup might help someone else go faster. That is a coherent safety argument. It is also an argument that makes the public version of the assessment structurally unable to do the thing assessments are for when the stakes are compounding: let outsiders check the work. Amodei's pacing essay asked the industry to take the rate seriously. METR measured a rate, redacted the dials, and left the rest of us with a number, a confidence interval, a co-founder's acceptance that checking just got harder, and a promise that Anthropic's own analysis is coming in the weeks ahead. Until that analysis arrives with more of the working shown, the sentence that stays with me is still METR's. The public must rely on their judgment more than they consider ideal. That is not a scandal. It is a design choice. Designs can be changed.