This article was drafted with AI (Claude) and reviewed by an editor. Both Anthropic and OpenAI appear here, and Anthropic makes Claude, so we have a stake. This is not an investment opinion on any company or stock; all figures are taken from reports and company announcements.
The biggest numbers story in AI this week was not performance but counting methods.
Case 1. Why one company's revenue went from $70 billion to $50 billion
On October 8, Axios reported that OpenAI's annualized revenue is $50 billion, not the $70 billion widely reported last month. The $70 billion came from material shared with investors and was "grossed up" so it could be compared more directly with Anthropic. The difference is how revenue from sales through cloud partners is recorded. According to an accounting professor cited by Axios, when a customer pays 100 through a cloud provider, Anthropic can record the full 100 as revenue and list the provider's cut as an expense, while OpenAI records only its own share on some partner sales. Both are said to comply with GAAP.
Reuters reported the same on October 9: OpenAI told investors its September annualized revenue was just under $50 billion. The gap from the earlier signal of nearly $70 billion came from an attempt to compare directly with Anthropic, and unlike Anthropic, OpenAI does not count revenue from sales through cloud partners such as AWS and Google Cloud (first reported by the Financial Times).
Annualized revenue stretches one period's sales over a year, a Silicon Valley-style metric, and Axios adds that public-market investors generally prefer actual revenue (all per Axios and Reuters; we have not seen either company's accounting records).
Case 2. The price per token can fall while the price per job does not
Anthropic announced that Haiku 5.5 is about 75% cheaper on average than Haiku 4.5 (company figure). But Next.ink, citing an outside analysis firm, reported that the model answers at greater length, so the savings depend on the effort setting, and also noted that requests over 100,000 tokens cost five times more per token. The AI Weekly newsletter also covered companies moving to cheaper models because of cost pressure and advises measuring "cost per completed task" (that issue summarizes other outlets' reporting, and we did not check the original reports).
Three questions for reading a number
- What was counted? Gross or net; tokens, or one finished job.
- Who produced the number? The company itself, or an independent party; and if they differ, under what measurement conditions.
- What happens if conditions change? A score at maximum effort may differ at the default level, and long requests may carry a different unit price.
Our take
Fact: One company's annualized revenue was reported as $70 billion and $50 billion depending on how it was counted, and this week it was pointed out that a drop in token price and a drop in cost per job are not necessarily the same.
Interpretation: Numbers can be compared only when units and counting methods match. That holds for comparing companies and equally for reading our own figures.
What we would do: In our revenue reports, separate gross from net, and record AI work as cost per "article, video or order" rather than per token. When citing another company's numbers, always state the source and the counting method.
Read together
We opened five sources of different kinds directly.
- axios.comOpenAI annualized revenue $20 billion less than previously reported (Axios)English, news report
- zawya.comOpenAI's September annualized revenue nears $50 bln, less than previously indicated (Reuters, via Zawya)English, wire report
- anthropic.comIntroducing Claude Haiku 5.5 (Anthropic)English, official company announcement
- next.inkChez Anthropic, Haiku 5.5 prend la relève des petits modèles rapides (Next.ink)French, outlet analysis
- aiweekly.coAI got too expensive, so companies are moving to cheaper models (AI Weekly #535)English, newsletter