Anthropic Files Its IPO as AI Agents Become Real Software

Anthropic's S-1 filing is the story of the day: the first real public financials test for a frontier AI lab. Coverage flags a "$441B problem" between compute commitments and revenue, and calls the filing "an AI doomsday primer" on capex risk (Fortune). The filing went up on sec.gov and instantly dominated Hacker News - 1,697 points, 1,659 comments (HN).

The filing also reveals the business behind the model: Claude-powered drug-discovery partnerships with Bristol Myers Squibb, Generate Biomedicines and BigHat, plus a $63B Broadcom compute financing (AI Affairs).

The math isn't the only question - the proof is

On the same day, OpenAI handed AI-generated math proofs to peer review: a Lean-formalized disproof of the Exponential Time Hypothesis for 3-SAT, a Kakeya/Stein restriction smoothing result, and a Yang–Mills proof (TheNextWeb). If they survive review, it's the clearest "AI does real research" milestone yet - and HN was dissecting it alongside the S-1 (HN).

Meanwhile, the agent layer is becoming actual product

• DeepSeek V4-Flash-0731 shipped on Hugging Face: a 284B-total/13B-active MoE, 1M context, MIT-licensed, re-post-trained for agentic work (HF).
• Off Grid AI launched on Product Hunt - private LLMs running fully on your phone and computer, no cloud (PH).
• 1Password for Agents is the quiet infra bet of the month: agents that sign in without ever seeing your passwords (PH).
• And the counterweight: data showing AI coding agents generate more code, but not more delivered software (Ars Technica).

Signal to watch

Whether OpenAI's formalized proofs survive peer review, and whether Anthropic's IPO numbers hold up under scrutiny - the two tests that decide if frontier-lab economics and frontier-lab research are real.