# METR finds AI roughly doubling Anthropic's research pace as its IPO filing flags a threshold

An outside evaluator spent eleven weeks inside Anthropic and concluded that AI is now speeding up the company's own AI research. Days earlier, Anthropic told IPO investors it may imminently cross a safety threshold for exactly that. What the numbers mean and how they are measured.

For years, "AI that builds better AI" was a thought experiment. In the first week of November 2026 it became a measured claim. On 5 November METR, a Berkeley-based non-profit that evaluates frontier AI systems, published an assessment based on about eleven weeks of high-access work inside Anthropic. Its central estimate is that Anthropic now improves its models on well-specified tasks at roughly twice its March 2024 to October 2025 pace, and that most of the speed-up probably comes from Anthropic's own heavy use of AI in research.

Three days earlier, Anthropic amended its IPO filing. The new text says the company "may imminently meet" its Responsible Scaling Policy definition of "dramatically accelerated progress attributable to AI", while no current model has been designated as exceeding it.

Together these are the first serious outside measurement of whether AI is accelerating AI development, and the first time a frontier lab has told investors it is near a safety tripwire for it. This article explains what "AI R&D acceleration" means, how it is measured, why labs set thresholds for it, what redaction does to trust, and what the same questions mean for engineering teams that use AI to ship software faster.

## What happened

### Anthropic amends its IPO filing (2 November 2026)

Anthropic filed its S-1 registration statement on 26 October 2026 and an amended S-1/A on 2 November. According to The Gadgeteer and a passage comparison by Emergent Mind, the amendment adds to the first risk factor that "our internal estimates and our expectations regarding future models, rather than our current models, indicate that we may imminently meet our Responsible Scaling Policy's definition of dramatically accelerated progress attributable to AI."

It continues: "We have not designated any of our current models as exceeding this threshold." Anthropic says it is "evaluating the implications" and considering "updates to our Responsible Scaling Policy to keep humans in control of frontier model development at Anthropic." The Gadgeteer also reports that the October text said Anthropic "may not identify" significant acceleration until after it occurred, while the November text says it has "in the past not identified" it until afterwards, turning a forecast into an admission. The same report says internal use of Anthropic's models, valued at list prices, grew more than 250 times between January and September 2026. That is a usage figure, not a measure of research progress.

### METR publishes its assessment of AI R&D acceleration at Anthropic (5 November 2026)

METR ran the evaluation from 12 August to 29 October 2026. Its summary, posted on the AI Alignment Forum, says Anthropic shared data on pretraining, reinforcement learning, training compute, internal benchmark scores, its own speed-up analyses and forecasts, and access to staff for interviews and surveys. The main findings:

- **Rate.** Anthropic is "likely improving model capabilities on well-specified tasks" at about 2x the historical rate, but "we would not be surprised if the true acceleration was closer to 1.3x or 3x."
- **Benchmarks.** Since October 2025, growth in AECI, Anthropic's internal version of the Epoch Capabilities Index, has run at about 1.9x the earlier rate (95% confidence interval 1.5x to 2.8x), or "roughly two years of historical progress over the last year."
- **Cause.** "A majority of this acceleration today is likely attributable to Anthropic's extensive use of AI to do R&D." Summaries of the report by AI/TLDR and by Karthik Tadepalli on LessWrong give METR's central guess as about 80%, with substantial uncertainty.
- **Compute.** METR is confident the acceleration "was not due to an acceleration in the scaling of training compute."
- **Assistance, not autonomy.** Research is accelerated "primarily via AI assistance rather than fully autonomous contributions", with signs that autonomous contributions will grow.
- **Bottlenecks.** None significant expected over the next 12 months, though METR is "not confident" because its view changed several times.

Fast Company reported that METR compared Anthropic's public August 2026 risk report with a private version and found the public one "overall understates the level of acceleration that was occurring."

### Much of the evidence is redacted

METR writes that "some of the key findings that most informed our conclusions about risk were redacted by Anthropic." Summaries say this covered most of the evidence METR found most informative, including per-stage estimates, compute breakdowns, bottleneck analyses and forecasts. The stated aim, as quoted in those summaries, was to "remove confidential information that could help other AI developers." AI/TLDR reports METR worked with about 2.5 full-time staff, on Anthropic devices, without query access to unreleased models.

## What "AI R&D acceleration" means

Picture a bakery that has always produced 100 new recipes a year. It hires an assistant who drafts recipes and runs test bakes overnight, and next year it produces 200 of similar quality. Asking whether the assistant accelerated the bakery means asking three things. Did output really double? Was it the assistant, or did the bakery also buy a second oven? Will it continue, or will the bakery run short of flour, tasters or shelf space?

For an AI lab the same three questions are **rate** (how fast capabilities improve against a baseline), **attribution** (how much is AI help rather than compute, staff, money or luck) and **trajectory** (whether bottlenecks cap the speed-up or it compounds). The safety concern lives in the third. If better AI makes research faster and faster research makes better AI, the loop can compound, a pattern usually called recursive self-improvement. METR's worry is a "sudden emergence of extreme and risky capabilities" that leaves everyone, the lab included, less time to react.

Be precise about the target. Anthropic's policy change log, describing version 3.1 from April 2026, defines the threshold as compressing "two years of 2018 – 2024 AI progress into a single year", and says this means doubling progress "in aggregate AI capabilities", not researcher productivity. Those two can differ a great deal.

## How you measure it

No single instrument measures AI-driven acceleration. METR combined several, and knowing their blind spots is the key to reading any such claim, including one about your own team.

### Capability indices: is the curve bending?

Epoch AI's Capabilities Index stitches results from many benchmarks onto one scale, so models tested on different benchmark sets can still be compared. Anthropic's internal AECI also covers unreleased models. METR fitted trends with a break in early October 2025 and compared slopes. By Tadepalli's reading, progress went from about 15 AECI points a year to about 29, which is the 1.9x figure.

Two cautions. Benchmarks favour tasks that can be scored automatically, and Fast Company reports METR's own caveat that the speed-up may not apply to "messy" work. Trend fits are also sensitive to the break point and the models included; AI/TLDR reports METR's robustness checks ranged from 1.1x to 2.9x. Tadepalli reproduced a speed-up from public data alone, rising from about 1.4x to about 1.9x as more models were added, which suggests the break is real rather than an artefact of private data.

### Task time horizons: how long a job can AI finish?

In March 2025 METR introduced the 50% time horizon: "the time humans typically take to complete tasks that AI models can complete with 50% success rate." It found this had doubled roughly every seven months for six years. The measure fits research work well. A model that reliably finishes ten-minute chores helps a researcher; one that finishes week-long projects replaces parts of a team. "Assistance rather than autonomy" is, in effect, a statement about where on this scale the help sits today.

### Uplift studies: does AI actually make people faster?

An uplift study compares people working with and without AI, ideally in a randomised trial. METR ran one in 2025: 16 experienced open-source developers each worked on about 15 issues from their own repositories, with each issue randomly assigned to allow or forbid AI. With AI allowed, tasks took 19% longer. Developers had expected a 24% saving and afterwards still believed they had saved about 20%.

That is the central lesson: **felt speed-up and measured speed-up can point in opposite directions**, like a driver on side streets who feels faster because the car keeps moving while the motorway would have been quicker. At Anthropic, Tadepalli notes, staff polls in October averaged a self-reported 12x uplift, which he converts with a simple model into roughly 1.6x research speed-up. METR calls such self-reports "challenging to interpret". The gap is not a contradiction: faster coding does not speed up training runs, deciding what to try next or reviewing results.

### Attribution: AI help or something else?

Showing that progress accelerated is easier than showing why. METR worked through the alternatives:

| Possible cause | How it can be checked | Reported view |
|---|---|---|
| More training compute | Compute growth before and after the break | Confidently ruled out by METR |
| More staff | Headcount trend | Steady, per Tadepalli |
| More data spending | Share of R&D budget | Too small to explain it, per Tadepalli |
| Earlier progress unusually slow | Compare with other labs and public trends | A reason the figure could be overstated |
| AI assistance | Timing, interviews, surveys, internal analyses | Likely the majority, about 80% central guess |

This is [causal inference](/tools/causal-inference) without an experiment: there is no second Anthropic without Claude, so diagnosis is by exclusion, as a doctor rules out causes. Tadepalli adds that the break coincides with Claude Opus 4.5 becoming available internally, but notes that the Opus 5.5 system card found a one-time jump fitted better than a lasting trend break. Settling that depends on models the public has not seen.

## Why labs write thresholds for this

A threshold in a safety policy is like a smoke alarm: installed in advance, silent most of the time, with an agreed response when it sounds. Writing it early means the decision is made calmly, not under deadline and commercial pressure.

| | Anthropic Responsible Scaling Policy | OpenAI Preparedness Framework v2 (April 2025) |
|---|---|---|
| Tracked | AI-driven acceleration of AI progress | "AI Self-improvement" |
| Main line | Two years of 2018 to 2024 progress in one year (v3.1 wording) | High: like giving every researcher "a highly performant mid-career research engineer assistant" versus 2024 |
| Higher line | Earlier versions named fully automating entry-level AI research | Critical: "recursively self improving", e.g. a generational improvement in "1/5th the wall-clock time" of 2024, sustained for months |
| Response | Risk reports, safeguards, possible external review; Anthropic says it may pause even where not required | High: stronger security controls. Critical: "halt further development" until safeguards exist |

Note the difference. OpenAI's High level is defined by researcher uplift; Anthropic's is defined by capability output. Since those diverge, a lab could be past one line and short of the other. The policy page lists version 3.4, in effect since 8 July 2026, as revising the automated R&D threshold "to better track the threat model of concern"; the public page does not give the revised wording.

The hardest part is timing. A capability trend is a lagging indicator: you only see the slope after several models exist, so a doubling may be months old before it is statistically clear. That is why the filing speaks of "expectations regarding future models", and why METR also examined forecasts and bottlenecks.

## Third-party evaluation and redaction

A company grading its own risk has a conflict of interest, the reason firms hire outside auditors. Fast Company reports that Dario Amodei's 12 September essay proposed external evaluators embedded in leading labs and named METR, and that Sam Altman also backs independent evaluators with employee-like access. METR accepts no payment, though partners provide access and model usage. Anthropic's policy log adds that public risk reports must mark where material was removed, unredacted versions go to at least 200 employees, and the Long-Term Benefit Trust can request external review.

Redaction creates two audiences. METR saw the evidence; the public sees conclusions. Ryan Greenblatt of METR wrote that conclusions beyond the benchmark rate "depend significantly on trust" and that "less extensive redactions would greatly reduce this need for trust." Ajeya Cotra was "sad that so much of the underlying evidence in this report was redacted." The commercial case is real, since a map of what sped up a lab is a recipe for rivals. The cost can be reduced by marking each redaction, publishing methods so others can reproduce what they can from public data, and giving at least one reviewer full access. As [a score is only as good as its report](/articles/a-score-is-only-as-good-as-its-report) argues, a number is only as credible as the evidence behind it.

## What it means in practice

**The "AI builds AI" loop is now operational.** An evaluator with internal access says it is probably happening at one lab, at roughly 2x, mainly through assistance. The debate has moved to what happens at the line.

**Checkable work speeds up first.** AI help is strongest where success is verified automatically: tests pass, a metric rises. Measure only that kind of work and you overestimate total speed-up.

**Felt is not measured.** A 12x self-reported uplift beside roughly 2x measured acceleration, and the 19% slowdown in the 2025 trial, are the same effect. Review, integration and waiting do not shrink when coding does.

**Disclosure is now a legal matter.** The warning sits in an SEC risk factor, where misstatements carry liability.

**Your team runs the same experiment.** If you use [Claude Code](/tools/claude-code) or similar agents, rate, attribution and bottlenecks apply to you, with the same measurement traps.

## What to do now

1. **Measure output, not feelings.** Track lead time from ticket to production, change failure rate and time to restore. Use surveys only as a supplement.
2. **Freeze a baseline.** Pick a pre-adoption window, as METR used March 2024 to October 2025, and do not move it after seeing results.
3. **Run a small controlled comparison.** Randomly assign comparable tickets to AI-allowed and AI-forbidden for a few weeks. Basic [statistics](/tools/statistics) such as confidence intervals stop you over-reading a noisy fortnight.
4. **Attribute before celebrating.** List what else changed (headcount, scope, tooling, a quiet quarter) and rule each out.
5. **Find the bottleneck.** If AI halves writing time but review takes 70% of the cycle, invest in review.
6. **Split checkable from judgement work.** Report speed-up separately for each. Good [evals](/tools/evals) and [regression boards](/tools/regression-boards) widen what you can safely delegate.
7. **Set your thresholds in advance.** Decide what level of autonomous AI contribution, such as merging without review or changing infrastructure, triggers extra controls, and keep a [human in the loop](/tools/human-in-the-loop) above it.
8. **Tag AI-produced work.** Label commits and experiments mostly done by AI, or you cannot answer the attribution question later.
9. **Publish method with numbers.** If you redact, say what and why, and give one independent reviewer the full picture. This is basic [AI governance](/tools/ai-governance).

## Risks and open questions

**The headline rests on one index.** AECI is an internal benchmark aggregate. Benchmarks saturate and favour verifiable tasks; METR lists both as reasons the figure could be too high.

**Attribution is judgement.** The roughly 80% share is a central guess with wide uncertainty, built partly on hidden evidence, and METR and the Opus 5.5 system card disagree on trend break versus one-off jump.

**Bottleneck analysis was unstable.** Hard-to-verify work such as research taste could cap acceleration at a new steady rate rather than a runaway one. The report cannot tell which.

**Most key evidence is redacted.** Readers must rely on METR's judgement, and METR's access fell short of staff level.

**"Crossing" is undefined in practice.** The filing does not say what changes once a model is designated over the line.

**One lab is not the industry.** No comparable outside assessment exists for other frontier labs.

**Watch next:** whether Anthropic designates a model as over the threshold and publishes the risk report; policy changes to keep humans in control of model development; whether other labs invite embedded evaluators; and any less redacted release of the evidence.

## Frequently asked questions

### What is AI R&D acceleration?

It is how much AI systems speed up the research that produces the next AI systems. Anthropic measures it by the rate of capability progress against a historical baseline, with doubling as the key line. The worry is a feedback loop that shortens warning time.

### Did METR say Claude is improving itself autonomously?

No. METR said the speed-up comes "primarily via AI assistance rather than fully autonomous contributions." People still direct the research, though METR saw signs autonomous contributions may grow.

### Has Anthropic crossed its threshold?

Not by its own designation. The 2 November filing says no current model has been designated as exceeding it, but that expectations about future models indicate it "may imminently" be met.

### Why not just ask researchers how much faster they are?

Self-reports are unreliable. In METR's 2025 trial developers believed AI made them about 20% faster when it made them 19% slower. Output data and controlled comparisons are more trustworthy.

### Why was so much of the METR report redacted?

Anthropic asked for redactions to keep out information that could help other AI developers. The cost is that outsiders must trust METR's judgement instead of checking the evidence.

## Summary

In early November 2026, AI-accelerated AI research moved from speculation to measurement. After about eleven weeks inside Anthropic, METR estimated that the company improves its models on well-specified tasks at roughly twice its earlier rate, with benchmark growth at 1.9x (95% interval 1.5x to 2.8x), training compute ruled out and most of the speed-up likely due to AI assistance. It judged that Anthropic's public August risk report understated the acceleration, while much of its own evidence was redacted. Anthropic's amended IPO filing says it may imminently meet its policy threshold, though no current model has been designated as exceeding it. For any team using AI, the lessons carry over: fix a baseline, measure output with controlled comparisons rather than feelings, attribute carefully, find the bottleneck and set thresholds before you need them.

## Sources

- [METR via AI Alignment Forum: Independent assessment of AI R&D acceleration at Anthropic](https://alignmentforum.org/posts/N5AmYkRks5CqAxvBg/metr-independent-assessment-of-ai-r-and-d-acceleration-at)
- [Fast Company: Outside researchers say Anthropic is speeding up its AI development](https://www.fastcompany.com/91636791/anthropic-metr-ai-claude-acceleration)
- [Let's Data Science: METR estimates Anthropic's AI capability progress has roughly doubled](https://letsdatascience.com/news/metr-estimates-anthropics-ai-capability-progress-has-roughly-e64b77bb)
- [AI/TLDR: METR on Anthropic, AI may have doubled the pace of Claude's progress](https://ai-tldr.dev/releases/metr-anthropic-ai-rd-acceleration/)
- [LessWrong: Anthropic's AI acceleration is probably real](https://lesswrong.com/posts/QgcgoMz7mviHL6Esr/anthropic-s-ai-acceleration-is-probably-real)
- [The Gadgeteer: Anthropic tells IPO investors its AI-builds-AI threshold may be imminent](https://the-gadgeteer.com/2026/11/03/anthropic-tells-ipo-investors-its-ai-builds-ai-threshold-may-be-imminent/)
- [Emergent Mind: Anthropic S-1/A, filed November 2, 2026](https://www.emergentmind.com/anthropic-threshold)
- [Anthropic: Responsible Scaling Policy](https://anthropic.com/rsp)
- [OpenAI: Preparedness Framework, Version 2](https://cdn.openai.com/pdf/18a02b5d-6b67-4cec-ab64-68cdfbddebcd/preparedness-framework-v2.pdf)
- [METR via forum mirror: Measuring AI ability to complete long tasks](https://forum.nunosempere.com/posts/YJ7Pk2bwTd3ieimG8/metr-measuring-ai-ability-to-complete-long-tasks)
- [Simon Willison: METR study on AI and experienced open-source developers](https://simonwillison.net/2025/Jul/12/ai-open-source-productivity/)