Reasoning token
Tokens spent by a reasoning model (o-series, DeepSeek R1, Claude with extended thinking) on hidden chain-of-thought before the visible answer. Billed separately at the same rate as output tokens. Can be 10-100× the visible-answer length on hard problems.
Background
Reasoning models thinkbefore they answer. The thinking happens in the same context window but is usually not surfaced to the user. Costs add up fast — a single math-olympiad problem can burn 50,000 reasoning tokens. APIs expose a reasoning-effort parameter (low/medium/high) so you can trade thinking depth against cost and latency.
Tools that use it
- 01→Claude
Anthropic's AI assistant with the Claude family of models (Opus, Sonnet, Haiku), Projects, and Computer Use.
- 02→ChatGPT
OpenAI's flagship AI assistant — GPT-5 chat with web, voice, image generation, code interpreter, and custom GPTs.
- 03→DeepSeek Chat
DeepSeek's web chat for DeepSeek-V3 and reasoning models, with open-weights releases.