Anthropic Warns That It Is Very Close to RSI Capability, Aka Self-Building AI

Anthropic says future models may soon meet its recursive self-improvement threshold, while METR estimates its benchmark progress has accelerated to 1.9 times

Updated November 20, 2026 at 2:38 PM ET

Quick Answer

Anthropic, PBC says future models may soon meet its recursive self-improvement threshold, but no current model has crossed it. The company also uses AI in research and development and is considering policy updates intended to preserve human supervision over frontier model development.

METR, an independent evaluator, found that Anthropic’s benchmark scores were growing at an estimated 1.9 times the pre-October 2025 rate by October 2026. That assessment provides new measurements behind the warning, although it does not establish that recursive self-improvement has arrived.

What Anthropic Disclosed About RSI

Anthropic filed an amended Form S-1 with the Securities and Exchange Commission on November 2, 2026. In the filing, the company said its internal estimates and expectations for future models indicate that it may imminently meet the Responsible Scaling Policy definition of “dramatically accelerated progress attributable to AI.”

The disclosure applies to future models rather than current ones. Anthropic said it had not designated any current model as exceeding the threshold and continued to assess the underlying data.

Anthropic also disclosed that it already uses AI in research and development. The company said AI-assisted development had produced models with capability gains it did not recognize at the time.

The filing separates AI-assisted research from fully autonomous development. Anthropic said it did not anticipate safely using Claude to design and build successive generations of frontier models without meaningful human supervision or understanding, if that becomes technically possible. It said it was considering updates to its Responsible Scaling Policy to keep humans in control of frontier model development.

What the Independent Evidence Shows

METR conducted a high-access evaluation of Anthropic’s AI research and development process from August 12 to October 29, 2026. Anthropic gave METR access to internal metrics covering pretraining, reinforcement learning, training compute, internal model benchmarks, AI speedup analyses, forecasts for a future model and interviews with relevant staff.

METR concluded that Anthropic was likely improving model capabilities on well-specified tasks at roughly twice its March 2024 to October 2025 rate as of October 2026. METR said the estimate carried high uncertainty and that the true acceleration could have been closer to 1.3 times or 3 times the historical rate.

METR also measured growth through Anthropic’s version of the Epoch Capabilities Index. The organization reported that benchmark scores had been growing at an estimated 1.9 times the pre-October 2025 rate since October 2025, with a 95% confidence interval of 1.5 to 2.8 times. According to METR, the result amounted to roughly two years of historical progress over one year.

METR attributed a majority of the current acceleration to Anthropic’s extensive use of AI in research and development, while noting significant uncertainty and other possible causes for acceleration in late 2025. The organization was confident that the acceleration did not result from faster scaling of training compute.

What Does This Mean for Business Visibility and Trust?

Anthropic’s filing does not report that a current model has reached recursive self-improvement. Its warning concerns future models and its continuing assessment of model capability data. METR’s measurements document faster capability progress, but they do not establish a self-directed development loop or show that current models have crossed Anthropic’s policy threshold.

For businesses, the practical advice is to keep the business name, contact details, address and service descriptions consistent across relevant profiles.

BizScoreAI can be used to track how a business listing appears in AI search results.

What to Do Now

  1. Compare the business name, contact details, website, hours, address and service description across the website and important directory listings.
  2. Record the authoritative page for each business detail.
  3. Correct outdated records and remove duplicates that present conflicting information.
  4. Save dated screenshots or exports of each corrected profile.
  5. Check the same details after a material change to the website, contact information, services or a directory profile.

Key Considerations

  • Anthropic’s threshold statement concerns future models; no current model has been designated as exceeding it.
  • METR’s 1.9-times estimate has a 95% confidence interval of 1.5 to 2.8 times.
  • METR said its broader rate estimate carried high uncertainty.
  • Anthropic is considering Responsible Scaling Policy updates intended to preserve human control over frontier model development.

FAQ

Has Anthropic Already Reached Recursive Self-Improvement?

No. Anthropic says none of its current models have crossed the recursive self-improvement threshold defined in its Responsible Scaling Policy, and it has not designated any current model as exceeding that threshold.

Does Anthropic Use AI to Build Its Models?

Yes. Anthropic says it uses AI in research and development and has used AI to develop models whose capability gains it did not recognize at the time.

How Fast Are Anthropic’s Model Capabilities Improving?

METR estimated that Anthropic was improving model capabilities on well-specified tasks at roughly twice its March 2024 to October 2025 rate as of October 2026. METR said the estimate carried high uncertainty.

What Did METR Measure?

METR measured Anthropic’s benchmark scores through the Anthropic Epoch Capabilities Index. It estimated that scores had grown at 1.9 times the pre-October 2025 rate since October 2025, with a 95% confidence interval of 1.5 to 2.8 times.

Will Anthropic Keep Humans Involved in Frontier Model Development?

Anthropic said it expects to keep humans involved in research and development for the foreseeable future. It does not anticipate safely using Claude to design and build successive generations of frontier models without meaningful human supervision or understanding, if that becomes technically possible.

Try the business directory

BizScoreAI, which includes the business directory

BizScoreAI scores how visible a business is to AI search and shows what its listing looks like to the engines people ask. Open the business directory.

Sources


See our editorial disclaimer for how our articles are produced.

🤖
Is your business visible to AI assistants?

Run a free scan to see your AI Visibility Score, SEO rating, and local citation accuracy.

Check Your Score →