OpenAI details its cost and infrastructure strategy

Sarah Friar details OpenAI's method for evaluating the true cost of a project, integrating processing time and generation errors.

Sarah Friar presents OpenAI's economic approach around three elements: model improvement, reducing execution costs, and increasing computing capabilities. The company believes that the cost of usage should not be evaluated solely based on the number of tokens, but also according to the time, errors, retries, and human intervention required to achieve a result.

This publication reviews the recent price reductions for GPT-5.6 Luna and Terra, as well as the Fast mode of GPT-5.6 Sol. OpenAI claims that several optimizations of its systems have reduced Sol's serving costs by 20% and improved generation efficiency by more than 15%.

The company also emphasizes the importance of coordinating the different layers of its business, from infrastructure to ChatGPT, Codex, and the API. Usage data from these products is used, in particular, to anticipate demand and guide future investments.

OpenAI reports having more than one billion active users and over two million enterprise customers. According to its internal data, users send about 50% more messages six months after signing up and use ChatGPT for twice as many tasks. Within the enterprise, agentic use cases via Codex reportedly represent 99.8% of weekly output tokens.

Finally, Sarah Friar specifies that future capabilities can be built directly, purchased, or developed with partners. Investment decisions must depend on usage growth, revenue, infrastructure utilization, and observed technical progress.