Anthropic S-1: In a few years, AI will autonomously do weeks' worth of work, and 90% of the code is written by Claude

Anthropic submitted its S-1 to the SEC on October 26. In addition to the offering prospectus, the filing also contains many specific predictions about AI capabilities. After reading it, several numbers are more worth remembering than most industry reports.

The filing's definition of Powerful AI is clear: within a few years, AI will reach or exceed top human experts in most professional and scientific fields; it will be able to work autonomously for hours, days, or even weeks; proactively seek clarification and feedback when needed; coordinate multiple instances and collaborate with humans and other AIs; and also control laboratory equipment, robots, and other machines through digital interfaces.

Anthropic calls this form a country of geniuses in a data center: millions of virtual agents working in parallel, covering fields such as software, physics, mathematics, biology, medicine, finance, and engineering. The company believes that exponential growth in cognitive capabilities will continue.

The S-1 devotes considerable space to closing the loop—using AI to accelerate the next generation of AI. Claude is already involved in experiment design, training data generation, automated evaluation, and security auditing. The most intuitive numbers: in September 2026, more than 90% of contributions to Anthropic's codebase were written by Claude. Each engineer's average daily code contribution was more than 45 times the 2024 annual average. TTM revenue per employee rose from $1.4 million to $6.6 million. This content has been audited and reviewed by lawyers; it is not a blog.

The financial data also provides some reference points: revenue was $41 million in 2023, $386 million in 2024, $4.585 billion in 2025, and $16.286 billion in the first half of 2026, up 1,300% year over year. Adjusted operating profit in the first half was $1.618 billion, and operating loss narrowed to $693 million. The net loss figure is large, mainly due to fair value changes in convertible notes, not day-to-day operations.

The filing also mentions that Claude discovered, in scientific research, molecular machines that could become a new direction for gene-editing tools; this is still early-stage. Anthropic believes AI has the potential to compress decades of scientific progress into a few years. At the same time, compute is still seen as a major constraint: more capable models will continually unlock new tasks, and demand will continue to exceed available compute. It also acknowledges that the economy has not yet made much use of many of today's model capabilities.

These statements appear in an IPO filing. In discussions about the filing on X, some users reminded others: the document's function is to sell stock, not to be an engineering blog. Some developers also said that their Claude Code starts to drift after running for a few hours and needs checkpoints; autonomous work for several weeks is still hard to believe. Another common question is: if it really runs for three weeks, who checks what it did in the third week?

The full filing is on the SEC website: https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm