Reasoning model

Reasoning models emerged publicly with OpenAI's o1 in late 2024 and became an industry standard within a year: Anthropic shipped extended thinking, DeepSeek released the open R1, and most labs converged on hybrid models where thinking can be switched on or its effort level tuned per request.

The mechanism is inference-time scaling: the model generates an internal chain of thought — often thousands of tokens — before producing the visible answer. Accuracy on hard math, science, and coding problems improves substantially, at the cost of slower responses.

Billing is the catch: thinking tokens are charged as output tokens even when hidden or summarized, so the same question can cost several times more with reasoning enabled. Effort controls and thinking budgets exist precisely so buyers can pay for deliberation only where the task warrants it.