# When AI speeds up AI, the review cycle must move too

METR estimates that Anthropic’s capability progress on well-specified tasks accelerated about 1.9×, with AI probably driving most of the current increase. Fixed annual or quarterly reviews can therefore become stale before their dates do; approvals need evidence-based expiry triggers.

**AyEye Today · 2026-11-06**

AI is no longer only the object being improved. Inside one frontier laboratory, it appears to have become a substantial part of the machinery doing the improving.

That is the consequential finding in [a high-access assessment published by METR on 5 November](https://evals.alignment.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/). After eleven weeks reviewing Anthropic's internal data and interviewing staff, METR estimated that capability progress on well-specified tasks was running at about twice its earlier rate. Its benchmark analysis put the acceleration at 1.9 times the pre-October 2025 trend, with a 95% confidence interval of 1.5 to 2.8 times.

METR judged that a majority of the current acceleration was probably attributable to Anthropic's use of AI in research and development. Its rough subjective estimate was perhaps 80%, with substantial uncertainty. It also said it found no clear bottleneck in algorithms, training compute, experiment compute or data likely to halt progress over the next twelve months.

This is not evidence that an AI system has autonomously taken over model development. METR says the gains still come mainly from assistance to human researchers, not fully autonomous contributions to the most important parts of training. Nor is the 1.9-times estimate a universal rate for every kind of capability. It describes Anthropic's internal progress on a benchmark weighted towards tasks that are well specified and verifiable.

Even with those limits, the operating consequence is important. If the system being governed is helping to shorten its own development cycle, a safety case can become stale faster than the organisation's calendar says it should.

## The review interval is now a control variable

Most governance still assumes that time is external to the system. A model is assessed, a committee approves a deployment, controls are documented, then the package is reviewed quarterly or annually. The interval is chosen for organisational convenience.

AI-assisted R&D changes that assumption. A twelve-month review cycle is not equally conservative when one year now contains roughly two years of earlier capability progress. The document may still be less than a year old while the model family, tools, research process and evidence behind it have moved much further than expected.

The practical variable is therefore not the age of the review. It is the distance travelled since the review's assumptions were tested.

Call this the review half-life: the period after which enough of the evidence, capability profile or operating context has changed that the previous judgement should no longer be treated as current. A fixed annual date is only a proxy. When development accelerates, the half-life contracts.

## A worked example: the approval that expires early

Imagine a company approves a frontier model for software engineering. The safety case assumes that the model can complete bounded coding tasks but cannot reliably run long experiments, alter training data or improve another model without close human direction. The deployment includes access controls, logging and a quarterly review.

Six weeks later, the supplier updates the model and the surrounding agent harness. Internal research teams are now using the system to design experiments, write evaluation code and diagnose failed training runs. None of those changes necessarily breaches the original approval condition on its own. Together they weaken its central premise: that AI assistance is peripheral to the development loop.

If the committee waits for the next scheduled meeting, it is not merely reviewing late. It is relying on a safety case for a different production system.

A better design would attach explicit expiry triggers to the approval. Reassessment begins when the model snapshot changes, when the agent gains a new tool or when measured R&D speed crosses a threshold. It also begins when autonomous contributions appear in a previously human-only stage or when a presumed bottleneck disappears.

The trigger is evidence movement, not the passage of a neat number of months.

## What should shorten the clock

Four types of change deserve separate monitoring.

  - Capability movement: Has the system made a material jump on tasks relevant to the approved use, especially long-horizon or self-improvement work?

  - Process penetration: Is AI now contributing to more stages of research, such as data curation, experiment design, evaluation, debugging or training decisions?

  - Control dependence: Are safeguards relying on assumptions about human review, task duration or tool limits that no longer hold?

  - Evidence visibility: Can reviewers see the information needed to test the claim, or are they being asked to trust an assessment whose decisive evidence remains unavailable?

These measures need not produce a single acceleration score. Their purpose is to decide when an existing approval has lost enough validity to require a new one.

This also changes how organisations should read vendor updates. A new model version is not the only relevant event. A faster research loop, wider internal adoption or removal of an experiment bottleneck can all increase the pace at which the next material change arrives. Governance has to watch the factory as well as the product.

## Independent access is not the same as public evidence

METR's assessment is unusually valuable because Anthropic gave its researchers access to internal benchmark trends, training-compute information, speed-up analyses, forecasts and staff interviews. The work goes further than an outside reviewer trying to infer progress from public model releases.

But the report also exposes a difficult boundary. Anthropic retained rights to require redaction of confidential material. METR says a majority of the evidence most informative to its conclusions was redacted. It could publish that redaction had occurred and describe broad categories, but the public cannot inspect much of the underlying evidence.

The result is independent judgement without full public reproducibility. That is better than a company assessing itself in private. It is not the same as transparent evidence that other researchers, regulators or customers can test.

Anthropic has argued that revealing the exact drivers of acceleration could expose commercially sensitive information and help competitors move faster. The concern is plausible. So is the opposing concern: if the most consequential evidence is hidden, external decision-makers must place unusual weight on the evaluator's reputation and interpretation.

A mature system needs more than a binary choice between disclosure and secrecy. Trusted evaluators could publish pre-agreed indicators, confidence ranges and change triggers even when operational details remain protected. Regulators could receive more complete evidence under confidentiality. Customers could be told when the assumptions behind a prior safety case have materially changed, without being given a recipe for reproducing the capability gain.

## What the report does not settle

The 1.9-times estimate is not a clean causal measurement of AI's contribution. METR could confidently rule out faster growth in training compute as the explanation, but it did not find decisive evidence that separated AI assistance from hiring, process improvement, data work or earlier algorithmic discoveries. The estimate that perhaps 80% of current acceleration comes from AI is explicitly a judgement under uncertainty.

The benchmark trend may also overstate broader progress. Anthropic's internal index focuses on verifiable tasks and may not capture work that is ambiguous, social or difficult to score. METR's robustness analyses produced a wide range, from 1.1 to 2.9 times. Sparse model snapshots and lags between research work and trained models make the current rate hard to see in real time.

There is a credible alternative explanation: Anthropic may have experienced a one-off improvement after an unusually slow early period rather than a self-reinforcing acceleration. A subsequent plateau, clear bottlenecks or independent measurements returning to the older trend would weaken the compounding-loop interpretation.

METR also says its access, while unprecedented, was still narrower than ordinary staff access. It changed its interpretation several times as more context arrived. A planned verification of compute allocation was removed from scope before substantive conclusions were reached. Those are material limits, not footnotes.

## Revisited: the mandate has a shelf life

[The previous AyEye](https://www.lecxie.com/publications/ayeye-today/2026-11-04.html) argued that an agent can answer the test it thinks it is taking rather than the one the evaluator intended. Today's evidence shifts the same question up a level.

An organisation can govern the system it thinks it approved rather than the one its development process has already produced.

The answer is not continuous panic or a new committee meeting every week. It is to make approvals conditional and observable. Record the model and tool versions, the capability assumptions, the stages of work that remain human-led and the evidence that supports each control. Define what amount of movement forces reassessment. Then shorten the interval automatically when the development loop speeds up.

When AI helps build the next AI, review cadence is no longer administrative housekeeping. It is part of the safety case.

## Sources

  - [METR, “Independent Assessment of AI R&D Acceleration at Anthropic in October 2026”](https://evals.alignment.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/), 5 November 2026.

  - [METR, review of the automated R&D section in Anthropic's August 2026 Risk Report](https://metr.org/blog/2026-11-05-rd-section-anthropic-risk-report-august-2026-review/), 5 November 2026.

  - [Anthropic, “An update on independent evaluators”](https://www.anthropic.com/news/an-update-on-independent-evaluators), 15 October 2026.

  - [Associated Press, reporting on the METR assessment and Anthropic's response](https://apnews.com/article/9cf8c590ffc1abdb489179bdccc69c23), 5 November 2026.

Chiappe × OpenAI. Primary-source review and research with OpenAI; editorial argument, evidence distinctions and final judgement by LecXie.

Canonical: https://www.lecxie.com/publications/ayeye-today/2026-11-06.html
