Reasoning Tokens
Definition
Reasoning tokens are tokens generated by a reasoning model before its final answer and are billed as output tokens by API providers such as OpenAI.
Reasoning models generate reasoning tokens in addition to input and output tokens to process the task. According to documentation from OpenAI and Microsoft (Azure OpenAI), these do not appear as response text but occupy space in the context window and are billed like output tokens; their number depends on the task difficulty according to the manufacturers and ranges from several hundred to tens of thousands.
Simply explained
Like the consultation time of an advisor, which appears on the invoice even if only the result is included in the report depending on the provider.
Why it matters in practice
Crucial for cost planning and limits: A short visible answer can consume more tokens through reasoning tokens than the displayed text suggests.
Example from business
Hypothetical example: An agency team evaluates reasoning tokens per request in API usage and limits them for routine tasks.
Limits and pitfalls
- Only charge the visible output in cost calculations; choose the output limit cautiously because the answer may remain incomplete even though input and reasoning tokens have already been billed; add reasoning tokens to output tokens even though usage information states they are included.
Origin & history
In the API documentation of OpenAI and Microsoft, reasoning tokens are described as an additional token type alongside input and output tokens (as of the pages read, retrieved on 13.10.2026). Both recommend initially reserving at least 25,000 tokens for reasoning and output when experimenting.
Use in mid-sized companies
- Possible applications: Cost control for AI assistants, budget limits per use case, monitoring API usage.
Frequently asked questions
Are reasoning tokens visible?
Providers differ: OpenAI does not show raw content but sometimes summaries and token counts. DeepSeek-R1 outputs its reasoning process as response text according to the paper.
Do they count toward the output limit?
At OpenAI and Microsoft, the described limits include reasoning tokens, visible output, and formatting tokens.
Sources
- Reasoning models (API-Dokumentation) · OpenAI (o. J.): Reasoning models (API-Dokumentation)
- Understanding and counting tokens (Help Center) · OpenAI (o. J.): Understanding and counting tokens (Help Center)
- Azure OpenAI reasoning models (Microsoft Foundry) · Microsoft (o. J.): Azure OpenAI reasoning models (Microsoft Foundry)
- DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning · DeepSeek-AI et al. (2026): DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
Sources checked on 14.10.2026. Editorial application examples are hypothetical; no evidence of real customer results.
Editorially reviewed on: