Anthropic released Claude Opus 5.5 on 22 September with a 230-page system card, and concluded that the model does not cross the automated AI R&D capability threshold in its Responsible Scaling Policy. That threshold turns on a specific quantity. Under version 3.4 of the policy, in force since 8 July, it is met if AI has doubled the rate of progress in aggregate AI capabilities, measured against both the rate Anthropic would expect and the fastest extended rate it has observed without significant AI contributions to AI R&D. The policy has been explicit since version 3.1 that this means capability progress rather than researcher productivity. The company's answer this month is no: its better-fitting reading of the trend has the slope unchanged after a one-time jump, and even the alternative reading puts the slope increase at roughly 1.5x against a line drawn at 2x. METR, the nonprofit that ran the external check, reached a compatible conclusion. What deserves attention is how much of the measurement underneath both answers has thinned out.
The threshold is a slope
Anthropic's policy gives two routes to the threshold. The first is substitution: whether its models could fully replace its entire set of research scientists and engineers at costs within a factor of five. The second is dramatic acceleration of the pace of AI progress for reasons likely related to the automation of AI R&D.
Substitution is not close. On an internal benchmark the card calls CoBench 2.1, Opus 5.5 scores within noise of Claude Mythos 5.1 and Claude Opus 5, well below the 85% figure Anthropic thinks a full substitute for its research staff would reach.
The acceleration route is where the measurement work happens. Anthropic tracks its rate of capability progress with a metric it calls the Anthropic ECI, described in the card as a fork of Epoch AI's Epoch Capabilities Index and powered partly by internal benchmark results, so its scores are not directly comparable to Epoch's public leaderboard. The question becomes arithmetic on a line: has the slope of AECI points per year doubled?
Two readings of the same curve
Anthropic fits a historical trend through eight releases from Claude Opus 3 to Opus 4.6, rising 14.75 AECI points per year. Opus 5.5 lands at 169.36, about 6.5 points above that line, and the card tests two explanations for the gap.
The first is a one-time jump of 5.9 points at Claude Mythos Preview, with the slope afterwards unchanged. The second is a genuine trend break, the slope rising from 14.4 to 22.2 points per year with the breakpoint fitted to September 2025. That second reading is a 1.53x increase in slope, with a 95% range of 1.20 to 1.82.
The one-time-jump explanation fits better, in 99 of 100 resampled fits. Anthropic's argument is that under either reading the slope has not doubled, so the threshold holds even on the less favourable interpretation. That is a sound way to structure the claim, though the upper bound of the trend-break estimate is 1.82 and the line is at 2.
Opus 5.5 is not what moved anything. It sits 1.24 points above Claude Mythos 5.1, and each model falls inside the other's local error bar; the card places the new model at or slightly above its previous frontier. The fitted acceleration, on the reading that finds any, dates to September 2025, a year before this model existed. Anyone arriving here expecting a runaway feedback loop should note that the document describes a shift that already happened and then stopped steepening, which is not the shape explosive-growth arguments require.
The index was rebuilt in the document that used it
For this card, Anthropic updated the basket of benchmarks making up the AECI and refit the index. The previous fit covered 338 benchmarks and 5,853 observations across 525 models; the new one covers 374 benchmarks and 7,985 observations across 732 models, plus corrections to inputs the company describes as previously incorrect.
Anthropic is direct about the consequence: because each refit re-estimates the whole scale, AECI values from this fit are not comparable to those in earlier cards. Recent frontier models rose by 4.1 to 6.1 points. Mythos 5.1, published at 162.0, becomes 168.12.
Re-fitting an index after finding errors in its inputs is the right thing to do, and none of it is concealed. It does mean the slope-doubling test was applied to an instrument re-specified in the same document reporting the result. The card also notes that scores far above its upper anchor, GPT-5 at 150, are less stable across refits, and frontier scores sit furthest from it.
A second disclosure matters more than it first appears. Anthropic writes that recent models have crossed the highest human baselines on many of the automated task-based AI R&D evaluations used in earlier system cards, that such results are no longer a significant component of its threshold determinations, and that it therefore did not run those evaluations for Opus 5.5. The suite that used to carry this decision has saturated and been retired from it. The August 2026 risk report had already flagged lower confidence than prior reports, citing that saturation and early signs of acceleration in one or more highly relevant internal metrics.
What METR could and could not see
METR's summary was published the same day and appears as the external-testing section of the card. METR's capability testing ran on API access granted over 10 business days across five tasks; the evaluation also drew on background trend data, Anthropic's answers to a questionnaire and an interview with an Anthropic researcher, and the separate METR report discussed below. METR states its work was not meant to verify compliance with any specific threshold in Anthropic's policies, and does not attempt to assess whether the model has particular alignment properties. The agreement was unpaid. METR drafted the initial summary, after which, in its own words, Anthropic had the opportunity to review and edit the text, and METR then signed off on the final version.
On capability the finding is unglamorous. Opus 5.5 improves on Claude Fable 5.1 across both verifiable and harder-to-verify tasks, but incrementally rather than as a discontinuous jump, and it retains qualitative weaknesses an expert human would be unlikely to exhibit on long-horizon or open-ended work. METR, while calling this highly uncertain, expects full automation of AI R&D to require large gains in foresight, prediction, building one's own feedback loops and other skills it says might typically be called researcher judgement or taste, and says its evidence does not suggest Opus 5.5 is a large improvement over Fable 5.1 in those skills. It also says its data cannot distinguish consistent from accelerating or decelerating rates of improvement, a direct constraint on what any of this settles about timelines.
The acceleration half is different in kind. METR's view that development was at least somewhat, but unlikely to have been dramatically, AI-accelerated rests on two grounds. One is its own reading that Opus 5.5 is an on-trend improvement over Fable 5.1, drawn from its capability tests and from Anthropic's questionnaire and interview answers, in which Anthropic says the model continues the AECI trend set from Mythos Preview onward. The other is a figure from a separate, highly experimental and preliminary METR assessment of AI R&D acceleration inside Anthropic, run by a different METR team with elevated access. That team shared its conclusions with the evaluating team but, METR writes, was not able to share the supporting evidence or the details of its reasoning. METR uses the figure as an input while stating that it does not argue directly in defence of the report's claims. It also drew on a further source it cannot currently disclose, which it says informed its view of AI R&D at Anthropic but did not relate to its understanding of this particular model.
The only acceleration figure the independent check cites is therefore a conclusion the checking team could not inspect, and its other ground draws partly on Anthropic's own account of the AECI trend. METR frames this as a trial of a more holistic process, with further public output expected in coming weeks. It is an honest attempt at a hard problem, and the kind of arrangement that third-party evaluation regimes and proposals for a dedicated standards body exist to make unnecessary.
Two numbers that both look like 1.5
There is a conflation here that is easy to make and wrong. The separate METR report's estimate is quoted as:
~1.5X overall acceleration in capabilities due to AI (i.e. 1.5 years in 1 year), with perhaps 30% chance of 2X acceleration.
Anthropic's trend-break reading gives a 1.53x increase in the AECI slope. These are not the same measurement. One is a ratio of progress to elapsed time attributed to AI; the other is a ratio between two fitted slopes of a capability index, breaking in September 2025. They land near the same figure, and treating that as corroboration would overstate what either document supports.
METR adds a caveat that narrows the 1.5x figure further: the preliminary report did not specify the time period the estimate covers, so it is unclear whether it applies to the development of Opus 5.5 or to some other stretch. The figure is already in the policy's own units, years of progress per calendar year, but without a stated period it cannot be tied to this model's development, and neither can the 30% chance of 2x attached to it. The two documents also use different comparators, Anthropic treating Mythos 5.1 as its previous frontier while METR benchmarks against Fable 5.1.
Where this rests on the company's own account
Substantial parts of the determination cannot be checked from outside. The internal measures of AI-driven research acceleration are, in Anthropic's phrase, only partially published; the card states they show no sustained AI-attributable 2x acceleration while acknowledging that some have moved and are being watched closely. CoBench 2.1 and the 85% substitution figure are internal. The judgment that the model is not close to replacing senior researchers comes from Anthropic's own pre-release use of it. The AECI is built partly on internal benchmarks.
METR's agreement includes a transparency provision letting it disclose, without Anthropic's consent, whether redaction rights were exercised and whether findings depended significantly on redacted information. The summary does not state either way, so on the public record it is unknown whether redaction occurred.
The defensible reading is narrow. A frontier developer has published a test of whether its own models have doubled its rate of AI capability progress, reported that it has not, shown its arithmetic, and had a third party examine part of it. That is more than the field had two years ago, and more than most scenario forecasts assumed would exist by now. It is also a test whose decisive instruments are internal, whose most concrete task-based evaluations have saturated out of use, and whose external check on the acceleration question rests partly on evidence the evaluating METR team was not shown.