Anthropic tells investors AI may soon speed up its research
The company has not designated any current model as crossing that bar, and it does not expect to safely let Claude build the next frontier models on its own without human supervision.
Jump to a section
Many people now open Claude or ChatGPT to draft an email, summarise a report or write a spreadsheet formula. Inside the companies that build these tools, AI has another job: helping researchers write code, run experiments and develop the next model. The better each new model gets, the more it can help with that work, and the question is how fast that cycle can speed up.
Anthropic, the company behind Claude, has just answered that question somewhere few people would look: a share sale filing. On 2 November 2026 it filed an amended IPO document with the US Securities and Exchange Commission (SEC), and TechCrunch found a passage in its risk factors saying that internal estimates about future models, not the ones in use today, suggest it may soon meet what its own safety policy calls "dramatically accelerated progress attributable to AI".
That does not mean it has happened. Anthropic said itself that it has not designated any current model as crossing the threshold, and that it is still assessing the data. All of this is the company's own forecast, not a finding checked by outsiders, and it sits in the same filing Anthropic made to sell shares and list on Nasdaq on 26 October, which you can read about here.
AI that helps build AI makes progress speed up in a chain
Picture how a new AI model gets built. Researchers design experiments, write code, wait for results and fix things, many times over. If AI can do more of that work, the new model arrives sooner, and if the new model is better than the last, it can build the next one faster still. People in the field call this chain, in which AI keeps speeding up the development of the next AI, recursive self-improvement, or RSI.
To give a sense of what "dramatically accelerated" means, Anthropic's Responsible Scaling Policy, the rules it set for which safety measures it must add as AI reaches each level of ability, once described its AI research threshold as AI doubling the pace of progress: squeezing two years of AI progress into a single year.
People in the field read that point very differently. Those who are worried see it as the moment AI could advance so fast that it slips out of human control. Enthusiasts see a big boost to work and scientific discovery. Anthropic also admitted in this filing that it has used AI to develop models whose leaps in ability it did not recognise at the time, and only realised later that development had sped up sharply.
A share sale filing must tell investors about the risks
So why is this in a share sale filing? A Form S-1 is the document a US company must file with the SEC before it sells shares to the public for the first time, known as an initial public offering (IPO). It must include a risk factors section, where the company tells investors what could hurt the business. An S-1/A is an amended version filed later, before the shares go on sale.
The risk Anthropic described is that the more it uses AI to speed up its own research, the faster its models' abilities could grow, possibly faster than its work on safety, on making AI act as people intend, and on security can keep up. Using AI to build more capable AI could also pose a considerable safety risk, it said. Put simply, the car could get faster before the brakes are ready.
Not everyone reads it that way. TechCrunch noted that some see an S-1 as both a disclosure document and a marketing tool, so a warning that the company is close to self-improving AI could also be promotion. At the same time, Claude has genuinely improved a lot this year in coding, maths and cybersecurity, which gives the warning some weight.
Anthropic does not plan to let Claude build its successors unsupervised
In the same filing, Anthropic said it does not expect to safely let Claude fully design and build successive generations of frontier models on its own without meaningful human supervision or understanding, even if that becomes technically possible. Frontier models here means the most capable models in the field at the time. The company is also considering updates to its Responsible Scaling Policy to keep humans in control of frontier model development.
That means in Anthropic's own plan, AI will help more and more with research but will not be the final decision maker. The company intends people to keep understanding and overseeing how new models are built. How that works in practice will depend on the next version of the policy, which it is still considering.
For Claude users, nothing changes this week
For people who use Claude or another AI tool at work every day, this passage does not change the apps they use this week, because it is about future models, not the ones available now. What could change later is the pace. If Anthropic's forecast holds, the tools we use could improve faster than before, and tasks people still do by hand today could get AI help sooner than expected.
For now, keep using AI as usual and keep checking its work before you use it. The thing to watch is the Responsible Scaling Policy, version 3.4, in force since 8 July 2026. If Anthropic updates the policy as its filing says it is considering, the new version will appear on the same policy page, which keeps a record of what each version changed.
FAQ
Can Claude improve itself now?
No. Anthropic said in its filing that it has not designated any current model as crossing this threshold. The warning is about future models, based on the company's internal estimates.
What is RSI?
RSI stands for recursive self-improvement: the point where AI is good enough to help develop the next AI, so that progress speeds up in a chain.
Why is this in a share sale filing?
A company selling shares to the public in the US must file a Form S-1 with the SEC and include a risk factors section for investors. Anthropic put this in that section of its amended filing of 2 November 2026.
Do Claude users need to do anything?
Not now. The apps and models in use have not changed because of this passage. Keep using them as usual and keep checking AI's work before you use it.
