Anthropic said in an amended IPO filing on 2 November 2026 that its estimates for future models indicate it may soon meet its own threshold for sharply faster AI research, TechCrunch reported. The company says no current model has crossed that threshold. The warning concerns the pace at which AI could help develop more capable models, with consequences for how quickly Anthropic can develop and check the safeguards for them.
Key points
- Anthropic says estimates concerning future models may imminently meet its threshold for dramatically accelerated AI research.
- The company has not designated a current model as exceeding the threshold and says it is still assessing the data.
- Anthropic does not expect to use Claude safely to build successive frontier models without meaningful human supervision.
- An outside evaluation judged Claude Opus 5.5 unlikely to fully automate AI research and development.
Anthropic’s forecast concerns future models
The amended filing follows Anthropic’s original IPO filing on 26 October. Its new language puts a nearer prospect alongside a risk the company had already identified: growing use of AI in its own research could make model capabilities advance faster than its safety, alignment and security measures. Anthropic says it has previously recognised a significant acceleration in development only after it occurred, and warns that this could happen again.
Anthropic bases its warning on internal estimates and expectations for future models, rather than a designation it has made for a model already in use. It says it is continuing to assess the underlying data. Those distinctions matter for a developer deciding what to build on Claude: the filing describes a possible change in the pace of Anthropic’s research, while the company’s stated concern is whether its precautions could keep up with such a change.
Recursive self-improvement can describe a much stronger prospect than faster AI-assisted research: a system designing and developing its own successor. Anthropic says it uses AI in research and development, but does not anticipate safely using Claude to design and build successive frontier models fully autonomously, without meaningful human supervision or understanding. A faster research cycle could still involve people choosing what to investigate, reviewing results and deciding what to build next.
Anthropic’s July policy sets the threshold
Anthropic’s Responsible Scaling Policy is its framework for managing risks from increasingly capable AI. Under version 3.4, effective 8 July 2026, one route to its threshold for dramatically accelerated progress requires an observed or expected doubling of overall capability progress attributable to AI, measured against both expected progress without AI and the fastest non-AI progress sustained across at least three model generations, Kingy AI reported. The policy has another route, involving AI taking over all its human scientists’ and engineers’ research duties at a price within fivefold of hiring people.
The acceleration test concerns the whole course of model development, rather than whether an assistant can complete one research task. It calls for a comparison with progress that would have occurred without AI assistance, as well as the earlier pace the company had managed without it. An improvement in one part of the work could leave other parts determining how quickly a new model is developed.
Writing and checking code could become quicker if a faster model-development cycle yielded better coding assistance. A doubling in research progress would not, by itself, measure time saved on either task: the policy’s comparison is about how quickly model capabilities advance across generations.
In an August risk report covering the period through 15 July, Anthropic judged AI-driven acceleration to be below its 2× threshold. It warned that AI handling research and model development on its own might pose a substantial concern in the next six to twelve months. The November filing places its estimate for future models closer to the policy threshold, while retaining the distinction between that estimate and its assessment of current models.
METR assessed Claude Opus 5.5 in September
METR, an independent organisation that evaluates AI systems, judged Claude Opus 5.5 unlikely to fully automate AI research and development in a September pre-deployment evaluation, Kingy AI reported. METR also judged it unlikely that AI had dramatically accelerated the model’s development. Its testing involved five tasks over ten business days, a bounded assessment of the model rather than a measurement of every part of Anthropic’s research operation.
A separate METR team with greater access made a preliminary estimate of roughly 1.5× capability acceleration and put the probability of 2× at about 30%, according to Kingy AI. The team responsible for the evaluation summary could not inspect the evidence behind that estimate, and the period it covered was unclear. Anthropic reviewed and edited the write-up, METR approved the final wording, and the assessment was unpaid.
Anthropic considers changes to human oversight
The filing puts the pace of safety work alongside the pace of model development as an investor risk. Anthropic warns that its increasing use of AI in research could make advances arrive faster than it can develop, validate and implement corresponding safeguards. AI Affairs has separately reported $895.4 billion in Anthropic computing commitments ahead of its IPO, another part of the company’s planned capacity to develop models.
Anthropic says it is considering updates to its Responsible Scaling Policy intended to keep humans in control of frontier-model development, TechCrunch reported.
.jpg){kind=link}