Watch the video, transcript and sources
+ Anthropic: public S-1, H1 2026 revenue $16.3B, net loss $80.7B (fair-value line $72.4B)
- LLM prices: cuts up to 80% (OpenAI), Opus 5.5 ~40% less to run
+ vLLM: constrained decoding 92% -> 100% tool-call schema accuracy
- Skin AI: 62% vs 42% on 200 lesions, with caveatsSixteen billion dollars of revenue in six months sounds like a company printing money. Anthropic's own prospectus lists an eighty billion dollar net loss for those same six months. The skin cancer study was drafted with a tool from the company in story one. I will come back to that at the end. A prospectus is the document a company files before it sells shares to the public. Think of it as release notes for a whole business, where every line has a lawyer's fingerprints on it. Anthropic, the company that makes Claude, filed its first one today.
Story one. Just after nine on Monday morning, New York time, Anthropic filed a Form S one with the S E C. It names the Nasdaq symbol A N T. It is a preliminary filing, so the share price, the share count and the newest quarter's revenue are still blank. Revenue was about forty one million dollars in twenty twenty three and about four point six billion in twenty twenty five. In the first six months of this year alone it was sixteen billion. Then the loss. The operating loss for those six months was about seven hundred million dollars, four percent of revenue. The net loss was eighty billion, and seventy two billion of that is one line, the change in fair value of convertible notes. In plain words, as the company becomes worth more, the notes it owes become worth more, and the accounts record that as a loss. The filing also reports an adjusted operating profit of one point six billion, which adds back stock pay and a few other items.
The same filing shows the copyright bill. In September twenty twenty five Anthropic agreed to settle the Bartz authors' class action for one point five billion dollars, which has its own line in the expense table. That settlement came up in our OpenAI pirated books story a month ago. Here is the part for developers. For the three months to the end of September, the filing says Claude Enterprise users who did not use Claude Code spent under seventy five dollars a month on average. The top five percent who did use it spent about twenty times that. The ownership is unusual. A trust elects a majority of the board through its own class of shares, and a company owned by the founders holds one special share worth fifty point one percent of the votes.
Story two. Max Woolf, who blogs as minimaxir, lined up every price change from OpenAI and Anthropic since June, and the post passed three hundred eighty points on Hacker News. OpenAI's own post says that on July thirtieth GPT five point six Luna got eighty percent cheaper. By September twenty ninth, OpenAI listed GPT six point one Sol at two dollars in and ten dollars out per million tokens, a fifth of the price of its top model, Astra. Anthropic says Opus five point five costs about forty percent less to run than Opus five, the price war we covered in September. Woolf's own reaction was to downgrade. He cut his Codex plan to a fifth of the usage limit, because the same model now stretched four times further. He also notes that in coding agents, cache reads are most of the token bill.
Anthropic's filing gives its own account. It says the September cuts combined a lower price with better efficiency, and that compute efficiency let it keep its margins. After the cheaper Opus four point five, it says, quarterly revenue growth went from forty eight percent to one hundred thirty two percent, with other factors helping. Woolf's explanation is open weight models. Kimi K three point one Pro, released October twenty third, lists at the same two dollars and ten dollars as GPT six point one Sol. So re-run your cost estimates against today's price pages. In agent work the cached input line matters most, and OpenAI cut that one in half for Sol.
Story three. Ubicloud, a cloud company, published a fix for tool calls in open models. Serving Kimi K two point six on vLLM zero point twenty eight, only ninety two percent of tool calls matched the tool's schema, which means two in every twenty five failed. The fix is constrained decoding. The server stops the model from picking any token that would break the tool's JSON schema. In vLLM zero point thirty one it is a flag called tool strict level, and Ubicloud measured one hundred percent on Kimi and DeepSeek. vLLM keeps it off by default because users reported problems with some models, such as GLM five point one. Ubicloud's point is that the engine and its settings change the quality you see.
Story four. Dermatologist Magnus Lynch, who works at Guy's Hospital in London, gave GPT six Astra and Claude Fable five point one a test of two hundred biopsied skin lesions, each as a close up plus a dermoscopy image. He also fine tuned Inception v three, an image network Google published in twenty fifteen, for about twenty minutes on one GPU. It was right on sixty two percent of the lesions. Each chatbot was right on forty two percent. The chatbots had favourites. GPT six Astra called seventy five lesions basal cell carcinoma when twenty five were. On melanoma, the one that matters most, the small network named thirteen of thirty first, against eight and eleven.
Lynch lists the caveats himself. Two hundred cases is small, each chatbot ran once, and the network is a baseline that regulated products beat. It runs in a phone browser through WebGPU and nothing is uploaded. Nothing here is medical advice. And the thread from the start. Lynch writes that the analysis and first drafts were produced with Claude Code, Anthropic's coding tool, and that one of the two chatbots he tested is also Anthropic's.
Verdict: NEEDS REVIEW — Big numbers, thin footnotes. Read the filing yourself.
Sources
https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm
https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/0001628280-26-067168-index.htm
https://news.ycombinator.com/item?id=50242939
https://minimaxir.com/2026/10/cheap-llms/
https://news.ycombinator.com/item?id=50236102
https://openai.com/index/advancing-the-price-performance-frontier-with-gpt-5-6/
https://openai.com/index/introducing-gpt-6-1-sol/
https://www.anthropic.com/claude-opus-5-5
https://openrouter.ai/moonshotai/kimi-k3.1-pro
https://artificialanalysis.ai/models/kimi-k3-1-pro
https://www.ubicloud.com/blog/fixing-tool-call-accuracy-in-vllm
https://lynchlab.net/llm-skin-cancer
And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.
YouTube · Newsletter · thedailydiff.dev · forward this to the intern who deployed on Friday.