Anthropic's S-1 shows compute under 50 cents per revenue dollar as OpenAI halves prices
The prospectus gives the first filed benchmark for a lab's unit economics. AI builders pay less per task, and public markets get a margin yardstick.
Competitive · Anthropic S-1 filed, Terra halves prices
Written for: Solo builder
Next step
Build the margin calculator using Terra's new price and your own product's tokens per task this week, then share it on Indie Hackers, where founders argue about pricing AI features.
Why now
The earlier card asked whether lab margins survive a price war. There is now a filed number to test, and in the same week OpenAI cut Terra's API prices by half. Commentators are already asking what happens if the IPO disappoints, and credit availability rather than chips is being called the bottleneck. For a solo builder the practical point is that token prices keep falling, so price your product per task or per outcome and keep the difference, instead of passing through token costs. This is news and context, not investment advice.
What became possible
Anthropic's S-1 shows compute and infrastructure spend falling from about $2.50 per $1 of revenue in Q1 2025 to under 50 cents in Q2 2026. It also lays out $800bn in data-centre commitments and an unusual governance structure, with about 100 of its 335 pages devoted to risks.
Two ways in
- Margin Calculator
A one-page tool: enter tokens per task and current price, and see your gross margin at Terra's new prices and at last month's, with a suggested per-task price.
- Customer
- Indie founders selling AI features
- Model
- Free tool that builds an email list for a paid pricing newsletter
- Lab Unit Economics Tracker
A public page that lines up each lab's price cuts against filed compute-to-revenue figures as more labs file.
- Customer
- Builders and readers who follow AI market stories
- Model
- Sponsorship or paid newsletter tier
Competition
The Lab Unit Economics Tracker is emptier, but the idea has a weakness. Price-history trackers already exist: Price Per Token, BenchLM, Epoch and open-source scrapers. Filing analysis is being done as one-off newsletter posts. Nobody yet puts the two side by side on one page that keeps getting updated, so that space is open. However, only one lab has filed, so the comparison has a single data point. There is also a problem with the "under 50 cents" number. The S-1 lumps training and inference together in operating expenses and gives no gross margin line. Analysts' estimates range from roughly 40% to 85%, so the yardstick is softer than the pitch suggests. The tracker could win by being the trusted, clearly sourced place that keeps these definitions straight and adds each new filing as it appears, for example from OpenAI. Until then it is mostly a content play, and it competes with established newsletters for readers.
Full competition notes (11 competitors)
The Margin Calculator idea is crowded. There are many free LLM cost calculators, and gross-margin guides for AI-wrapper founders are common. Adding a per-task price is a small step that any existing tool could copy in an afternoon. The real opening is timing: the Terra cut came out on October 26 and existing calculators still show June or July prices. A before-and-after view of the Terra price change, which shows margin and a suggested price instead of just cost, could get noticed this week. After that it will be hard to stand out unless it connects to real usage data, such as your billing or API logs, instead of numbers you type in. The Lab Unit Economics Tracker is emptier, but the idea has a weakness. Price-history trackers already exist: Price Per Token, BenchLM, Epoch and open-source scrapers. Filing analysis is being done as one-off newsletter posts. Nobody yet puts the two side by side on one page that keeps getting updated, so that space is open. However, only one lab has filed, so the comparison has a single data point. There is also a problem with the "under 50 cents" number. The S-1 lumps training and inference together in operating expenses and gives no gross margin line. Analysts' estimates range from roughly 40% to 85%, so the yardstick is softer than the pitch suggests. The tracker could win by being the trusted, clearly sourced place that keeps these definitions straight and adds each new filing as it appears, for example from OpenAI. Until then it is mostly a content play, and it competes with established newsletters for readers.
- Coworker LLM Cost Calculator
A free calculator that compares monthly cost across models using published per-token prices. Its prices were last updated in June 2026. It shows cost, not margin or a suggested price, and it does not compare old prices with new ones.
- CalcPlanet AI API Cost Calculator
A generic token-cost calculator. It covers the same basic math as the proposed Margin Calculator.
- Wavect LLM cost calculator 2026
An engineering guide and calculator for per-request LLM cost. It warns that its example prices go out of date.
- Beancount.io LLM API cost and AI wrapper gross margin guide
Content written for indie AI-wrapper founders about counting inference as cost of goods and working out gross margin. It targets the same founders the Margin Calculator would, but it is an article, not a tool.
- Price Per Token pricing history
Charts per-model price changes over time from OpenRouter data. This is half of what the Lab Unit Economics Tracker would show, but it has no data from company filings.
- BenchLM LLM pricing trends
Output price history by model and a price index starting March 2023. It tracks prices only, not lab economics.
- Epoch AI LLM inference price trends
A well-regarded research dataset on how inference prices fall over time. It is a likely source for the price side of a tracker, and a likely competitor too.
- llm-price-tracker (PyPI)
An open-source tool that saves copies of the official OpenAI, Anthropic, Google and other pricing pages and compares them over time. Anyone could use it to build a price-cut timeline.
- Kepler Anthropic S-1 Breakdown
Already pulls the compute-to-revenue numbers from the S-1: about 48% of revenue in Q2 2026 and about 55% for the first half. This is the filing-data side of the proposed tracker, as a one-off analysis.
- Clouded Judgement: Benchmarking the S-1 Data
One of many newsletter posts benchmarking Anthropic's S-1 against public software companies. Tunguz, SaaStr, OnlyCFO and others cover the same ground, so the margin-yardstick story is already well covered.
- The Neuron on GPT-6.1 Terra price cut
Confirms the trigger: Terra costs $1 per million input tokens and $5 per million output, down from $2 and $12. The news is fresh, and none of the calculators above has been updated for it yet.
Checked 2026-10-27
Risks
- The S-1 figures are chosen by the company and may not hold after the IPO
- The IPO may be delayed or priced lower, changing the story
- A credit squeeze could end the price war and raise token costs
Strong tech readiness and buildability; market pull and open space and team fit are the weak point (5).
tech 8 · pull 5 · space 5 · fit 5 · build 8
Exposed
Companies that could benefit or be threatened. News and context, not investment advice.
Momentum of its main topic
Sources
- S-1: Anthropic spent ~$2.50 on compute and infrastructure for every $1 of revenue in Q1 2025, falling to under 50 cents in Q2 2026, as it seeks profitability (Financial Times)Techmeme
- Anthropic S-1Hacker News
- Will the real Anthropic please stand upMIT Technology Review
- The Anthropic S-1 Filing Is Also an AI Doomsday PrimerWired AI
- GPT-6.1 Terra halves OpenAI's API prices and still gains ground on its larger modelsThe Decoder
- 4 Questions for AI This WeekThe AI Daily Brief
- AI:AM Highlights: Clean Neurons, Black October, the GPU Credit Crunch, and Buying the Supply ChainThe Cognitive Revolution
Status history
- 2026-10-27 Added to the watchlist