Anthropic and OpenAI do not publish what it costs them to serve a model. Everything below is reporting or analyst estimates, relayed through secondary articles we could read on 18 October 2026. We did not read the original reports from The Information, the Wall Street Journal or SemiAnalysis, so each claim carries the name of the outlet we did read. Nothing here is a statement about what any provider pays, and nothing here is a prediction.
What each source says
| Source and date | What it reports | Definition and caveats |
|---|---|---|
| WebProNews, 22 January 2026, relaying The Information | Anthropic cut its 2025 gross margin projection to 40 percent from 50 percent, because inference costs were 23 percent higher than it had anticipated | Gross margin defined as revenue minus inference costs and other product sales expenses. The article does not say whether the figure is actual or projected, and says it could not verify the underlying numbers |
| AI Afterhours (Substack), 15 April 2026, an opinion analysis of WSJ reported investor documents | Gross margin of 40 percent for Anthropic and 33 percent for OpenAI. The author says inference takes more than half of revenue for both, and that about a third of OpenAI's inference spend supports free users | Leaked, unaudited documents. The author notes definitions and accounting differ between the companies and that some figures are his own estimates |
| Crypto Briefing, 10 October 2026, relaying SemiAnalysis | Inference margins "could achieve margins as high as 88%". Roughly 75 to 85 percent of Anthropic's annual recurring revenue is API. Consumer subscriptions are about 10 percent of revenue but more than 40 percent of inference compute, at roughly 50 percent compute margins | An analyst estimate, not company data. The 88 percent is described as an upper bound for inference only, excluding model training |
The same SemiAnalysis coverage says overall gross margin rose from negative 94 percent in 2024 to the mid 60 percent range by mid 2026. WebProNews separately reports a 70 percent compute margin for OpenAI in late 2025, attributed to The Information via SaaStr.
Why the numbers do not agree
They measure different things. A "compute margin" counts revenue after direct inference. A "gross margin" in the first two sources also includes other cost of sales. An "inference margin" in the SemiAnalysis coverage is an estimate for serving only. A 40 percent gross margin and an 88 percent inference margin can both be reported without either being wrong, because they answer different questions.
They also cover different periods, from 2025 actuals or projections to estimates for mid 2026. And they come from different kinds of source: internal projections, leaked investor documents and analyst modeling.
What you can safely take from this
- The sources describe subscriptions and API use as having different compute shapes. SemiAnalysis, via Crypto Briefing, says subscribers get over 4 times the compute per dollar of API buyers, while the Substack author says free users take a share of spend. Both are claims about mix, not about your bill.
- None of the sources gives a per token cost that you can plug into a budget.
- Reported margins are about the vendor, not about whether a particular plan is a good deal for you.
Margin arithmetic, to read the numbers correctly
A margin is a share of revenue, not a multiple of cost. This snippet shows what a few reported margins imply for each dollar of revenue.
for margin in (0.33, 0.40, 0.50, 0.88):
cost_per_dollar = 1 - margin
print(f"{int(margin * 100)}% margin means cost of {cost_per_dollar:.2f} per $1.00 of revenue")
That is arithmetic on a definition. It does not say which cost lines each source included.
Why this matters for a buyer
Pricing decisions are better made on published prices and your own usage than on guesses about vendor costs. Our posts on what a Claude plan buys and how to estimate a session use only list prices and labelled scenarios for that reason. Safeguard lets you bring your own model, which keeps the choice of vendor yours.
Quick check
Why can a 40 percent gross margin and an 88 percent inference margin both be reported?
They measure different cost sets. One includes other cost of sales and the other covers inference only, and they come from different sources and periods.
Do these sources tell you what a token costs a provider?
No. They report margins and mix, not a per token cost.
