METR judges that AI is accelerating AI development: Anthropic is progressing 1.9x faster, 80% thanks to AI itself

METR finds AI accelerates AI development: Anthropic is progressing 1.9x faster, 80% thanks to AI itself

🔎 Why this report is a game-changer

On November 5, 2026, METR published what no one had measured from the outside before: the actual rate at which a cutting-edge lab improves its models using AI. Verdict: Anthropic is progressing approximately 1.9x faster between March 2024 and October 2025, and about 80% of this acceleration is attributable to the use of AI in its own R&D.

This is neither a marketing press release nor a theoretical extrapolation. It is an independent evaluation conducted with high-level access to the lab's confidential data over eleven weeks, from August 12 to October 29, 2026, including interviews, surveys, and unpublished model data.

For the first time, an external organization has quantified a self-acceleration loop forming. Fears of recursive self-improvement, long confined to speculation, now have a first measurable point—reported by Reuters, The Verge, and the Associated Press.

But read until the end before citing these figures. The majority of the most informative evidence was censored by Anthropic, uncertainty is high, and METR acknowledges that its analytical choices influence the result. This report is a historic milestone—and an exercise in methodological humility.

Key takeaways

• Pace multiplied by ~1.9: AECI scores (Anthropic's version of the Epoch Capabilities Index) are growing at 1.9x the pre-October 2025 pace [95% CI: 1.5–2.8x] — i.e., roughly two years of historical progress in one year.
• ~80% attributed to AI: the majority of the acceleration is probably due to extensive use of AI in R&D, with substantial uncertainty around this figure.
• Compute is not the cause: METR is confident that the acceleration is not coming from scaling of training compute.
• No bottleneck in sight: no marked diminishing returns have been observed, so there is a significant risk that progress accelerates further over the next 12 months.
• Redacted evidence: the majority of the most informative evidence has been redacted by Anthropic.
• Assistance, not autonomy: AI primarily accelerates things through assistance to researchers, not yet through major autonomous contributions.

Recommended tools

To verify for yourself, here are the resources to prioritize — all free.

What this report actually measures

Direct answer: METR measures the pace of progress in Anthropic's capabilities on well-specified tasks, then attempts to attribute that pace to its causes. It does not measure an AI that "improves all by itself."

The evaluation ran from August 12 to October 29, 2026 — eleven weeks of highly privileged access: confidential data, data on unreleased models, surveys and interviews with employees. It is this level of access that distinguishes this report from a mere outside analysis paper.

The central instrument is the AECI, Anthropic's version of the Epoch Capabilities Index: an aggregate of capability benchmarks. According to METR, these scores are growing at 1.9x the pace observed between March 2024 and October 2025, with a 95% confidence interval of 1.5x to 2.8x.

To give a sense of the public scale: in the leading agentic benchmarks, Claude Opus 4.7 (Adaptive) scores 94.3, behind GPT-5.5 (98.2) and Gemini 3 Pro Deep Think (95.4). These indices aggregate dozens of standardized tasks — exactly the type of metric METR leveraged, in internal versions and with more data.

An important point: METR judges that the pace of improvement is "probably" ~2x that of March 2024–October 2025, with values from 1.3x to 3x plausible. The 1.9x is the central estimate on the AECI index, not an absolute truth. As The Verge summarizes it, the doubled "probably" in the key sentence is no editorial accident: it reflects the actual state of the evidence.

The key figures, as published

Direct answer: four figures stand out from the report — and they don't all carry the same degree of reliability.

The 80% figure is an attribution estimate: METR is attempting to apportion the acceleration among its possible causes. This is the most fragile exercise in the report, since it relies largely on surveys and interviews — and thus on human self-reports.

The "12x" figures reported in Slack polls from October 2026 are individual perceptions of productivity. They are consistent with an overall doubling of the pace, but none of them is a controlled measurement.

By contrast, the conclusion about compute is solid: METR states it is confident that the acceleration does not stem from an acceleration in training compute scaling. This is precisely what makes the "AI helping to create AI" hypothesis credible.

A self-accelerating loop? What the report says — and doesn't say

Direct answer: the report does not demonstrate an autonomous self-improvement loop, but it provides the first external estimate compatible with its existence.

The key distinction is the one between assistance and autonomy. METR is explicit: AI accelerates research mainly through assistance to researchers, not yet through major autonomous contributions. The models suggest leads, write code, speed up experiments — but do not conduct R&D on their own.

That doesn't stop the dynamic from being concerning. Reuters coverage highlights that the study feeds fears of recursive self-improvement (RSI): a system that improves the system that created it. The fact that ~80% of the acceleration comes from AI itself, rather than from compute, matches the profile expected at the start of such a loop.

METR's third official conclusion: absent any observed bottleneck or marked diminishing returns, there is a significant chance that the acceleration will spiral over the next 12 months. This is a risk judgment, not a prediction — but it is the first time an independent evaluator has formulated it on real data.

This picture echoes what we wrote about Anthropic's call for a global AI pause, the 80% of code written by Claude, and self-improvement that is accelerating. The lab itself documents a growing dependence on its models — the METR report provides the external quantification of it.

What would be needed to confirm the loop

Three conditions, at minimum: uncensored data, replication at other labs, and a direct measurement of AI's autonomous contributions to R&D. None of the three is met today. The loop remains a strong hypothesis — but a hypothesis that now has a number.

Censoring the evidence: an audit that can't be audited

Direct answer: the majority of the most informative evidence was redacted by Anthropic, which means the public has to take METR at its word.

That's the central tension of the document. METR had access to confidential data and unreleased models, but the most compelling passages were redacted before publication. The outside reader cannot verify the raw data, the interviews, or the exact construction of the index.

Anthropic's defense comes down to one word: confidentiality. A lab can't publish data on unreleased models or its R&D internals without shooting itself in the commercial foot. It's the classic trade-off between trade secrets and public verifiability.

But the effect remains the same: we have an "independent" evaluation whose core evidence cannot be independently verified. This needs to be said clearly, because some of the press coverage repeats the 1.9x figure and the 80% without mentioning the redactions.

On the other side of the ledger: eleven weeks of access to internal data, with a published methodology and explicit caveats, is still exceptional for the industry. Most labs publish nothing comparable. The problem isn't what METR did — it's what the industry's standards of transparency allow.

Methodology: five reasons to read with a critical eye

Direct answer: METR itself acknowledges that its analytical choices affect the 1.9x estimate, and that the link between benchmark scores and real capabilities remains uncertain.

1. A benchmark index, not a measure of real capabilities

The AECI aggregates scores on standardized tasks. But a benchmark score measures performance on those tasks, not the ability to do real research. METR says so explicitly: this link remains uncertain. An index rising 1.9x faster does not guarantee models that are 1.9x more useful.

2. The choice of the reference period

The ratio compares October 2025–October 2026 to March 2024–October 2025. Changing the endpoints changes the ratio: if the reference period was atypically slow, the acceleration would appear stronger. METR publishes an interval (1.5–2.8x) precisely to reflect this sensitivity.

3. Attribution based on self-reports

The ~80% rests on surveys and interviews, with self-reported uplifts of up to 12x in Slack polls. These are individual perceptions, not controlled measurements. And METR notes the absence of consensus among Anthropic employees on the magnitude of AI's role.

4. An evaluation dependent on access granted by the company being evaluated

METR could only work with the access granted by Anthropic, which then redacted the evidence. The analyst's independence does not fully compensate for the dependence on access. No comparable public evaluation exists for any other lab, which precludes any comparison.

5. "Well-specified" tasks, not open-ended research

The conclusion applies to well-specified tasks. But real R&D includes a large share of ill-defined problems, which benchmarks capture poorly. The measured acceleration could overestimate — or underestimate — the gain on exploratory research.

The report also deserves credit where it is strong: the negative finding on compute is presented with high confidence, and the caveats are published in black and white. That's rare, and it's worth reading.

On the theoretical side, the science of representations is progressing too: the Platonic Representation Theorem by LeCun and Balestriero shows that models trained on different data converge toward similar representations. A useful foundation for understanding why aggregated capability indices make sense — and why they remain proxies.

The context: an acceleration under strain

Direct answer: this report arrives at the worst possible moment for Anthropic's caution narrative — and at the best possible moment for its finances.

The external evaluation is published amid calls to slow down, as reported by the Associated Press. Yet the lab in question is the very one that called for a global pause on AI, while documenting that 80% of its code is now written by Claude. Calling to slow down while progressing twice as fast: the tension is visible.

On the financial side, Anthropic amended its S-1 with $1.17 billion in Q3 revenue and $8.95B in compute commitments — a trajectory that assumes demand keeps pace with capabilities. An acceleration of progress supports that bet, but it also raises the pressure on announced safety timelines.

The METR report makes this dilemma measurable: the lab says it wants to slow down if necessary, but its own AI-boosted pace of progress makes that "if necessary" harder to define with each passing month. This is exactly the kind of situation that responsible deployment frameworks are meant to govern — and that this report puts to the test.

What This Means for Developers and Freelancers

Direct answer: in the short term, nothing changes in your tools; in the medium term, expect significantly more capable models, faster — and plan accordingly.

First implication: the pace. If the rate of progress doubles sustainably, the generation jumps—such as moving from Claude Opus 4.6 (84.7) to Claude Opus 4.7 (Adaptive) (94.3) in agentics—will become closer together. Technically fixed architectures will age faster.

Second implication: value shifts. When models become better faster, what remains rare is domain expertise and access to an audience. Laboratories are automating their R&D; professionals should automate their distribution.

This is where AI avatars make sense—provided you don't confuse the technologies: an AI avatar is not a chatbot, neither in architecture nor in the value produced. An avatar embodies expertise; a chatbot answers questions.

For experts who want to catch this wave rather than be swept away by it, the detailed strategy for monetizing your expertise through an AI avatar offers a concrete framework. The irony of the story: the same dynamics that accelerate cutting-edge labs make this lever accessible to freelancers.

❌ Common Mistakes

Mistake 1: Reading "1.9x" as an exact figure

It's a central estimate with a 95% confidence interval of 1.5x to 2.8x, and METR notes that values from 1.3x to 3x are plausible depending on analytical choices. Always cite the interval, never the point estimate alone.

Mistake 2: Believing "80%" means AI does 80% of the work

The 80% is an attribution of the acceleration, not a share of the work. It's a highly uncertain estimate, with no consensus among Anthropic employees. AI primarily accelerates work through assistance, not through major autonomous contributions.

Mistake 3: Confusing capability acceleration with compute scaling

METR is confident that the acceleration does not come from scaling up training compute. This is precisely what makes the result interesting: the cause would be the use of AI in R&D, not a bigger compute budget.

Mistake 4: Treating the "12x" figures from Slack polls as measurements

These uplifts, reported in polls from October 2026, are individual perceptions. They are consistent with an overall doubling of pace, but none of them is a controlled measurement. Never cite them as facts.

Mistake 5: Citing the report without mentioning the redactions

The majority of the most informative evidence was redacted by Anthropic. Quoting the 1.9x and the 80% without this caveat presents as publicly verified something that is not.

❓ Frequently Asked Questions

What is METR?

METR (Model Evaluation & Threat Research) is an independent organization specialized in evaluating the capabilities and risks of cutting-edge AI models. It is notably known for its work on the horizon of agentic tasks and its high-access evaluations conducted prior to the deployment of frontier models.

Does this report prove that AI improves on its own?

No. It shows that the use of AI in R&D likely accelerates progress (~80% of the acceleration), mainly through assistance to researchers. This is the first external estimate compatible with a self-acceleration loop, not proof of autonomous self-improvement.

Why did Anthropic censor the majority of the evidence?

The report does not publicly detail the reasons, but the censored elements concern confidential data and unpublished models—so likely industrial secrets. The price to pay: an evaluation whose main evidence cannot be verified externally.

What will happen if no bottleneck appears within a year?

METR judges that there is a significant risk that the acceleration will continue, or even run wild, over the next 12 months. Specifically: closer generations of models, safety thresholds reached faster, and increased pressure on laboratories' commitments to responsible deployment.

Where can I read the full report and reactions?

The full report is published on the METR website on November 5, 2026, along with its methodology and reservations. For a synthesis of press reactions and post-publication analyses, the briefing du 7 novembre is a good starting point.

✅ Conclusion

For the first time, an independent evaluator has quantified an AI self-acceleration loop — 1.9x, ~80% driven by the AI itself — but with evidence that is mostly redacted and an uncertainty that calls for caution. Read the full METR report and form your own judgment.