Anthropic says in its IPO filing that it may soon reach the threshold at which AI doubles the pace of its own development

Anthropic says in its IPO filing that it may soon reach the threshold at which AI doubles the pace of its own development

On November 2, Anthropic, the developer of Claude, filed Amendment No. 1 to its IPO registration statement with the U.S. Securities and Exchange Commission. The amendment adds a forecast about reaching the threshold, acknowledges that the company has previously been slow to recognize substantial acceleration, and describes a requirement it is considering for human review of certain model development decisions.

The original filing acknowledged that the company might recognize a substantial acceleration in its development only after it had occurred, because much of its data on the pace of development consists of lagging indicators. The amendment says this has already happened (“we have failed to notice in the past”) and adds a forecast:

“Our internal assessments and expectations regarding future models, rather than current ones, indicate that we may soon meet the definition of ‘dramatically accelerated AI-driven progress’ in our Responsible Scaling Policy.”

The company defined this threshold itself in its Responsible Scaling Policy. The threshold is reached if the company observes or expects AI progress to be twice as fast as anticipated and twice as fast as any previous sustained progress achieved without substantial AI assistance, with research automation plausibly accounting for much of that doubling. “Twice as fast” means “as much progress in one year as would normally be seen in two.” The threshold is intended to identify “the onset of dramatic recursive self-improvement,” a cycle in which each version of a model helps build the next.

According to the filing, Claude helps develop subsequent models by running experiments, preparing training data, automating model evaluations and safety audits, and writing and reviewing code. In September, Claude wrote over 90% of contributions to the codebase.

In its August risk report, the company assessed the pace as still below the threshold and said it might cross it “within the next year.” In September, METR, an organization that studies advanced AI, published a summary of its evaluation of Claude Opus 5.5 before release. A separate METR team provisionally estimated that AI had accelerated improvements in Anthropic’s model capabilities by a factor of approximately 1,5, with an approximately 30% probability that the pace had doubled. The summary itself concludes that AI was unlikely to have dramatically accelerated the development of Opus 5.5.

The company does not expect to be able to safely entrust Claude with fully autonomous design and development of subsequent model generations without meaningful human oversight or understanding, if and when this becomes technically possible. It is therefore considering updating its policy to require people to review and approve “certain decisions” in research and development. According to an article from Anthropic’s own institute, humans set the goals for engineering tasks and Claude finds ways to achieve them, but Claude remains substantially less capable than humans at choosing the goals. The institute describes this gap as the distance still to be covered before a system can independently design its successor.

The same article identifies human code review as a bottleneck: the overall pace is constrained by the parts of the process that have not accelerated. In biotechnology, experts interviewed for the August report identified animal testing and clinical trials as such constraints. Some said that AI had not substantially changed the path from a drug candidate to a patient, which typically takes 9–15 years. Even if computing resources are the only constraint on model development, the institute writes, “more intelligence cannot discover what a drug does over decades of use.”