Newshot Anthropic's model R&D accelerated twofold, but the public only sees a filtered version
METR, after an 11-week internal investigation at Anthropic, says its models completed about two years' worth of historical R&D progress in one year, and employees using the strongest internal models self-reported productivity of about 12 times that without AI; currently the acceleration mainly comes from AI-assisted human research, but Anthropic heavily redacted the roughly 100-page report.
- METR believes about two years of R&D progress was completed in one year
- Employees self-reported productivity reaching 12 times that without AI
- Three new models were trained consecutively within half a year
- The public report removed most of the most informative evidence
You see in AI news that "models covered two years of progress in one year," while at the same time Anthropic is preparing to go public and its CEO publicly advocates slowing down development. Put the two together, and what really needs judging is not the promotional slogan, but which contributions from AI the acceleration comes from, and how far the public report can actually prove.
One year covering two years: who backs this conclusion
METR spent 11 weeks inside Anthropic, and its conclusion is: over the past year, the company's model R&D covered about two years of historical progress. The researchers are also confident that this round of acceleration cannot mainly be attributed to scaling up training compute.
Benchmark scores are the first basis: the score curve of Anthropic's models climbed within one year the distance that previously took two years to cover. METR's wording is "about two years"—not that every model and every task is exactly twice as fast.
| Basis for judgment | What it shows |
|---|---|
| Benchmark scores of Anthropic's models | About two years of historical R&D progress completed within one year |
| METR internal survey | Employees self-reported that after using the strongest internal model, work efficiency was about 12 times that without AI, twice that when using Mythos 5 a few months earlier |
Compute expansion has long been the main driver of model progress, and this time METR ruled it out: the researchers are confident that the bulk of the acceleration is not from buying more chips or piling on more training time. As of October, their attribution was that most of the acceleration comes from AI itself, and this contribution is still growing over time. This judgment is supported by anonymous internal Slack polls and employee interviews, not by controlled experiments in a lab.
There is another layer that is easily overlooked: employees are using internal models that outsiders cannot access. Over the past half year, Anthropic finished training three unreleased models, codenamed Model 2, Model 3, and Model 4, completing training in July, August to September, and September to October respectively. That is to say, the 12x figure cannot be reproduced with the Claude that ordinary people have.
This evaluation has so far only been done by METR, and has not yet been externally replicated. "One year roughly equals two years" is not yet a law repeatedly verified by independent teams; it is a judgment with premises: the premises are that both the benchmark scores and employee self-reports are credible, and that the gap between internal models and external versions does not affect the direction of the conclusion.
The 12x figure was filled in by employees themselves
In METR's report, Anthropic employees self-reported productivity at 12 times that without AI assistance. The number comes from Slack questionnaires and interviews; it is individuals' estimates of their output, not a third party's re-measurement of R&D results.
This report comes from the AI safety research organization METR (an organization that evaluates AI R&D progress), and the investigation was conducted inside Anthropic for 11 weeks. The 12x figure was filled in by employees themselves in Slack (an internal corporate chat and polling tool), cross-checked with interviews, and points to a considerable increase in output capacity, rather than an independently verified R&D acceleration curve.
Looking at a slightly longer timeline, this self-reported level is about twice that when using Mythos 5 (the model Anthropic provided to employees at the time) a few months earlier. The same group of people, the same set of questionnaires, a stronger tool, and the self-reported output gap also doubled—this is the comparison in the report that best illustrates "acceleration is steepening," but both sides are self-reported measures, not actual measurements.
But 12x cannot be directly equated with a 12-fold shortening of the model training cycle, nor can it prove that AI can already independently set, verify, and advance R&D plans.
Right now it is AI assisting humans; autonomous contribution is only a trend. METR attributes most of the current acceleration to AI-assisted human research, not AI working independently; "autonomous contribution" here means AI completing part of the research tasks on its own without step-by-step human instructions.
The report notes that fully autonomous contributions may soon play a larger role—this is a trend judgment, not a fact that has already happened. Researchers are still responsible for posing questions, choosing methods, and judging results, while models help them advance specific work faster.
Three models trained in half a year, none released
Model 2 finished training in July, Model 3 in August to September, and Model 4 in September to October—three models squeezed into half a year, and not one has been released.
According to The Verge, METR (an independent organization evaluating AI R&D acceleration), after its internal investigation at Anthropic, gave these three unreleased models their codenames. "Unreleased" matters here: they are not models the public can already use in products, yet they have already entered Anthropic's research pipeline. METR's report points out that a large amount of the acceleration and productivity gains comes from these unreleased internal models.
Why can internal models become the pivot of acceleration? Researchers use stronger new models to complete the next stage of work, and the results then become training material for the next generation of models. Capability improvement therefore no longer depends only on scaling up training compute, but can, through models participating in R&D, hand the feedback from one round of experiments to the next round earlier.
METR also mentions an unreleased model traceable to 2022, Unreleased 2022 (an internal model traceable to 2022), indicating that models remaining in the internal R&D chain are not limited to this batch of new projects.
Anthropic consecutively completed training of three new models within half a year, forming an internal relay: unreleased models first boost researchers' production capacity, then send new results into the next round of training.
No ceiling in sight yet, but don't treat it as a guarantee
METR did not find a clear mechanism that would prevent further acceleration. This is an observation at the time of the investigation, not a promise about the future.
"Diminishing returns" here (more input, lower return) means that with each additional unit of resources invested, new results become fewer and fewer. METR did not observe clear signs of this, nor did it find an observable bottleneck (a sticking point that has already appeared and hinders the process from advancing). Not observing it does not mean it will not appear.
What is driving R&D right now is still human-machine collaboration, not an automated loop detached from humans. The report only says "there are already signs" that fully autonomous contributions (work AI completes without step-by-step human involvement) will soon account for a larger share.
The 100-page report had its most informative parts removed when made public
The full evaluation report METR submitted is about 100 pages, and the public version was redacted by Anthropic, removing most of the evidence the researchers considered most informative.
Under the agreement between the two parties, Anthropic has the right to redact the final report, and it used that right thoroughly. METR put it bluntly: "We submitted a full evaluation report of about 100 pages to Anthropic. In the public version, Anthropic removed most of the most informative evidence we found."
What was removed was not background material. METR says this evidence could have had multiple reasonable interpretations, and after its removal, the public cannot judge for itself whether the conclusion holds up and must rely more on METR's judgment. It also notes that its information channels are "relatively narrow," and a few data requests were denied.
Anthropic co-founder Sam McCandlish responded bluntly on X: "We removed a lot, and I accept that this makes it harder for others to check METR's work. But we judged that it is not appropriate to publicly disclose what exactly is driving the acceleration—this is both commercially sensitive and amounts to handing peers an acceleration roadmap." He said the company will publish its own report in the coming weeks.
The conclusion that one year covered two years is currently supported by only one redacted piece of public evidence—the verifiable space ends here, and the rest must wait for Anthropic's own report.
Sources: The Verge, "Why a new report on Anthropic automating AI development is alarming researchers" (November 7, 2026), and official materials from METR and Anthropic; figures are paraphrased according to the disclosure wording in the report and official pages, and no inferences are made about the redacted parts of the public report.
Glossary · Plain-language notes appearing in this piece
The full text of in-depth pieces, research interpretations, and tool tutorials, plus all podcasts, are in the member area.