Anthropic’s amended IPO filing puts useful numbers behind its warning that AI could accelerate the company’s research: as of August, Claude led 26% of research tasks, while the share completed fully autonomously remained zero. The figures sharpen the distinction between AI helping researchers and AI taking over research itself.
Anthropic Watch analysis
What happened
A detailed analysis by Trending Topics says Anthropic’s filing describes Claude leading a research task when it does the work end to end but an engineer decides whether the result ships. The share rose from under 1% in February to 26% in August. The analysis also says the filing puts fully autonomous research at zero in every area measured.
The article reports that more than 80% of Anthropic’s merged code lines now come from Claude, with the filing citing over 90% for September. Anthropic cautions that lines of code are an imperfect measure and that the real productivity gain is lower. A March poll of 130 researchers estimated a fourfold output uplift from Mythos Preview, though the article says Anthropic described the actual figure as “somewhat lower”.
Why it matters
These measures show why “AI helps build AI” can mean several different things. Claude can contribute heavily to coding and complete research work under human direction, while the reported measure of fully autonomous research remains at zero. Those are not contradictory claims; they describe different degrees of human involvement.
The numbers also make the company’s warning more concrete, without proving that a self-improving loop is already operating. The distinction between assistance, human-supervised work and autonomy is where the meaningful evidence sits, not in a dramatic label.
Our read
The most useful takeaway is not that Claude is about to build its successor unaided. It is that AI is already doing a growing share of work inside a frontier lab, while the reported autonomy measure remains at zero. That is a more interesting and testable story than either “nothing to see” or “the machines have taken the lab keys”.
What to watch
- Whether Anthropic publishes updated figures for research tasks led by Claude and tasks completed fully autonomously.
- How the company defines and measures human involvement in those tasks.
- Whether its stated productivity estimates translate into independently assessed research output.
Discussion spark: Should “AI-led research” count as meaningful autonomy when a human still decides whether the work ships, or is that distinction the whole point?
Sources and evidence
- Anthropic official source (4 November 2026, 15:46 UTC)
Anthropic Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Anthropic.