METR's Anthropic AI R&D acceleration assessment explained: what the 1.9x means, why it ranges from 1.1x to 2.9x, and what was redacted

METR spent 11 weeks inside Anthropic's data and concluded that AI-driven progress there now runs about 1.9 times its old pace. We explain the index behind that number, why one modeling assumption moves it from 1.1x to 2.9x, and what the public cannot check.

This explains reporting by

METR: Independent Assessment of AI R&D Acceleration at Anthropic in October 2026 (November 5, 2026). Read the original first:

https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/

In one minute #

• METR ran an 11-week review, August 12 to October 29, 2026, with about 2.5 full-time staff and data Anthropic chose to share.
• Its main number: Anthropic's capability trend has grown about 1.9x faster since October 2025, from about 15 to about 29 index points a year. The 95% interval is 1.5x to 2.8x.
• That number depends on a modeling choice. Swapping the link function, which assumes how hard tasks inside a benchmark are spread, gives anything from 1.1x to 2.9x, and the data fits most of them about equally well.
• METR's subjective estimate is that about 80% of the October 2026 speedup comes from AI itself. Staff Slack polls put their own speedup near 12x, which METR calls very loose evidence.
• Anthropic redacted most of the evidence METR found most useful, and cancelled one planned check midway. The public has to take much of this on METR's judgment, as METR itself says.

What METR looked at #

METR, an independent AI evaluation group, published the assessment on November 5, 2026. It asked three questions: has the growth of Anthropic's internal capabilities sped up, how much of that is due to AI, and are there bottlenecks that could slow it.

The review covered internal-only models Anthropic calls Model 2 (training finished July 2026), Model 3 (August to September) and Model 4 (September to October), plus public models such as Mythos Preview for comparison. METR did not get query access to unreleased models. It worked from Anthropic's internal benchmarks, about 15 staff interviews, headcount, code and pull request counts, training compute, inference spend and internal efficiency analyses. A full report of about 100 pages exists but was not published.

The index behind the 1.9x #

The main measure is AECI, Anthropic's version of the Epoch Capabilities Index. METR used Epoch's public code. The idea is close to how standardized tests are scored: each model gets one hidden ability score, and each benchmark gets a difficulty and a discriminability, meaning how sharply scores change as ability passes the difficulty. A model's expected score on a benchmark is a curve of discriminability times ability minus difficulty.

Fit that across many models and benchmarks and you get one ability number per model over time. METR then fit straight-line trends before and after a break fixed at October 7, 2025. Before the break the index rose about 15 points a year. After it, about 29. That ratio is the 1.9x, or roughly two years of the old progress in one.

A second fit in the post supports a lasting change of slope rather than a one-time jump at Mythos Preview: the slope-change model's error was 1.9 points against 3.9, and it won in 81% of bootstrap draws.

Why one assumption moves it from 1.1x to 2.9x #

The curve that turns ability into a score is called the link function. It encodes an assumption about how the tasks inside a benchmark are spread by difficulty: how many are trivially easy or broken, how many are nearly impossible. METR's default is the logistic curve, which assumes a symmetric spread. Here is what the other common choices give for the same data:

link function            Model 4 acceleration
complementary log-log    1.1x   (assumes many easy tasks)
probit                   1.8x
logistic (default)       1.9x
learned Stukel           2.6x
cauchit                  2.8x   (heavy tails on both sides)
log-log                  2.9x   (assumes many hard tasks)

The intuition: if a benchmark is assumed to hold many easy tasks, the last points of improvement near the top are cheap, so recent gains count for less and the speedup shrinks. If it is assumed to hold many hard tasks, the same late gains mean much more ability, so the speedup grows. METR reports that most of these curves fit the data about equally well out of sample, around 86 to 87% R squared, so the data cannot pick a winner.

Our reading: the fair one-line summary is not 1.9x. It is somewhere between about 1.1x and 2.9x, with METR's best guess near 2x. METR also reports that across all frontier developers, not just Anthropic, the same method gives about 1.4x, and that Anthropic's own Haiku 5.5 system card reported 1.7x, with the gap mostly from METR including internal-only models.

There is a ceiling problem too. Model 4 scores above 50% on all 15 benchmarks METR used. As benchmarks fill up, this kind of index gets harder to read, which METR flags itself.

Measured speed versus felt speed #

Anthropic staff have voted in Slack, with emoji, on how much the latest model speeds up their work. Those polls rose from about 1.2x for Claude 3.7 Sonnet in January 2025 to about 12x for Model 4 in October 2026. METR calls this very loose evidence, since the people who answered may not be typical and the answers took little effort.

The 12x and the 1.9x measure different things. One is how much faster a person feels on their tasks. The other is how fast the models' measured ability is rising. A person can feel many times faster while the lab's overall progress speeds up far less, because research also waits on compute, experiments and decisions.

The interviews sat in between. Coded by METR with Claude's help, they came out 2 yes, 3 leaning yes, 4 borderline, 0 leaning no and 2 no, which METR reads as mildly favoring a 2x or larger speedup.

Where the 80% comes from #

The share due to AI is a judgment, not a measurement. METR says most staff on the project shared an estimate of about 80% for October 2026, with large uncertainty. The evidence it cites: it is confident training compute did not grow faster, the interviews centered on about 2x, self-reported uplift keeps rising, there are early signs of AI making autonomous contributions, and non-AI factors would not explain a lasting speedup. It thinks non-AI factors drove more of the late-2025 change, but those details were redacted.

METR also says AI cannot yet make highly significant, fully autonomous contributions to the most important parts of training, and found no convincing bottleneck in algorithms, compute or data for the next 12 months, with low confidence.

What nobody outside can check #

• Anthropic had unilateral redaction rights and redacted most of the evidence METR found most informative, saying the aim was to remove information that could speed up other developers.
• The benchmark names are redacted. METR lists only the groups: math, software engineering, science, multimodal, and language or memory.
• A planned check of how compute is split between capabilities and safety was cancelled midway at Anthropic's request.
• METR marked four items of the AI Evaluator Forum's AEF-1 checklist as not met: query access was not requested, method details cannot be shared, redaction rights were not limited to concerning findings, and METR has no responsible disclosure policy.
• METR leaned heavily on Claude for data collection and analysis, and notes that could distort results. It judged deliberate misleading extremely unlikely.
• One team member had discussed a role at Anthropic and held equity in a competitor; an unconflicted member checked their work.

What to watch #

Anthropic's Responsible Scaling Policy lists AI R&D acceleration as a key risk indicator. METR says it did not judge any policy threshold and aimed for a whole picture instead. Its stated worries are a significant risk of further speedup over the next year, autonomous contributions arriving soon, and benchmarks running out of room to measure. A separate METR review of the AI R&D section of Anthropic's August 2026 risk report is being published alongside.

Who is affected #

What to do #

• When you see a 2x AI R&D headline, ask which link function and which models it rests on.
• Read METR's caveats section before quoting the 80%; it is a stated judgment, not a measured share.
• If you plan model upgrades on a yearly cycle, expect the gap between releases to keep growing in capability, and treat METR's range as the honest bound.
• Watch for METR's separate review of Anthropic's August 2026 risk report and any public version of the full report.

What is still unknown #

• The names of the 15 benchmarks and the internal results behind them are redacted.
• The roughly 100-page full report has not been published.
• The non-AI factors behind the late-2025 speedup are redacted.
• Whether the index tracks real-world impact is open; METR says crossing a capability threshold could matter more than the trend shows.
• No outside group has reproduced the numbers, and METR relied on Claude for parts of the analysis.
• Whether the trend holds past October 2026.

Sources #

AI News Report· every headline, every morning.