On November 2, 2026, Anthropic said in an amended Form S-1 that its estimates for future models suggest they may imminently meet the company’s threshold for “dramatically accelerated progress attributable to AI.” In that filing, Anthropic said it had not designated any current model as exceeding the threshold. The forecast concerns accelerated AI research and development, not a demonstrated system building successor models on its own. Anthropic’s S-1/A
For context on the wider debate, NeoTeo’s earlier coverage of researchers’ warnings about self-improving AI explored the broader risks. Anthropic’s filing focuses on a specific policy test for automated research and development.
What Anthropic’s threshold measures
Anthropic’s Responsible Scaling Policy (RSP) v3.4, effective July 8, 2026, sets out two alternative routes to its automated AI R&D threshold. One concerns AI systems taking over the work of the company’s research scientists and research engineers. The other concerns a sustained acceleration in aggregate AI capability progress. (RSP v3.4)
| RSP route | Requirement |
| Full substitution | AI systems could replace Anthropic’s entire set of research scientists and research engineers at competitive cost, defined as within a factor of five. |
| Accelerated progress | Aggregate AI capability progress is observed or expected to reach twice both the expected rate and the fastest extended rate seen without significant AI contributions. The acceleration must be plausibly and substantially attributable to automated AI R&D, assessed across at least three model generations. |
“Double the rate” means making as much progress in one year as the baseline would produce in two. The measure is aggregate AI capability progress, not a doubling of individual researchers’ productivity. Anthropic also said much of its evidence about AI-driven acceleration relies on lagging indicators and that it has sometimes recognized significant acceleration only after it occurred. (RSP v3.4; S-1/A)
Why the forecast is not a self-building loop
Recursive self-improvement broadly describes AI helping to develop more capable successor systems. Anthropic’s forecast concerns a policy threshold for accelerated R&D; it does not say that Claude is already designing and building successive frontier models without people.
In the November 2 filing, Anthropic said it did not anticipate being able to use Claude safely to design and build successive frontier-model generations fully autonomously without meaningful human supervision or understanding. The company also said it was considering updates to its RSP to keep humans in control. Those are separate from the accelerated-progress threshold, which the filing framed as a possible development involving future models.
