Are Reasoning (Thinking) Tokens Billed as Output Tokens?

Updated 2026-10-25. Prices from models.json, checked against official pricing pages on 2026-10-25.

Yes. OpenAI, Anthropic, and Google all bill a model's hidden reasoning ("thinking") tokens at the normal output-token rate. None of them has a separate, cheaper reasoning price. You also pay for the full reasoning, even when the API returns only a summary of it or nothing at all.

For agent builders this is often the biggest surprise on the bill. A short visible answer can carry thousands of output tokens you never see. Here's how each provider handles it, how to measure it, and what it costs.

How each provider bills reasoning

OpenAI

OpenAI's reasoning guide says reasoning tokens "are not visible via the API" but "still occupy space in the model's context window and are billed as output tokens." In the usage object they're included in output_tokens and itemized under output_tokens_details.reasoning_tokens:

"usage": {
  "input_tokens": 75,
  "output_tokens": 1186,
  "output_tokens_details": { "reasoning_tokens": 1024 }
}

In that example, 1,024 of the 1,186 billed output tokens are reasoning, and only 162 are visible answer.

Anthropic (Claude)

Anthropic's thinking docs say tokens Claude spends reasoning "are billed as output tokens, even when the thinking text isn't returned to you." Two details matter:

• You pay for the full thinking, not the summary. Claude returns a summary of its reasoning (display: "summarized") or an empty block (display: "omitted"). Either way you're charged for the full thinking tokens generated, not the summary tokens, so the billed output count won't match what you see.
• Earlier thinking can be billed again as input. On newer models (Opus 4.5 and later, Sonnet 4.6 and later, Haiku 5.5, Fable/Mythos 5 and 5.1), the API keeps thinking blocks from earlier assistant turns in context by default. Those kept blocks are billed as input on later requests, like the rest of the conversation.

Check usage.output_tokens_details.thinking_tokens to see how much of output_tokens was reasoning. output_tokens is the authoritative billed total.

Google (Gemini)

Google's pricing page labels the output row for its reasoning models as "Output price (including thinking tokens)." Thinking is simply part of output.

What output costs on popular reasoning models

Per 1M tokens, standard tier, short context. Prices are from TokenTab's models.json and were checked against the official pricing pages on 2026-10-25.

*Promotional prices. Gemini 4 Argon goes to $4.00/$20.00 on Dec 1, 2026. Gemini 3.8 Flash goes to $1.50/$7.50 on Jan 1, 2027.

On every model here, output costs about 5x input. That's why a burst of hidden reasoning hurts more than the same number of prompt tokens.

Worked example: one request

Say a request has 2,000 input tokens. The model thinks for 3,000 tokens and returns a 300-token answer. This mirrors the example OpenAI's community forum uses: 3,000 reasoning plus 300 visible equals about 3,300 billed output tokens.

On Claude Sonnet 5.5 or GPT-6.1 Sol ($2 in / $10 out):

• Input: 2,000 × $2/M = $0.004
• Output: 3,300 × $10/M = $0.033
• Total: $0.037

If you had only counted the visible 300-token answer, you'd have estimated $0.007. The real cost is about 5x higher, and reasoning makes up about 81% of the total.

Worked example: a 30-step agent

Agents reason on every step. Say each of 30 steps generates 2,000 reasoning tokens. That's 60,000 reasoning tokens per task, before any visible output:

Run that agent 1,000 times a month and reasoning alone costs $600 on Sonnet 5.5 or $3,000 on GPT-6 Astra. If earlier thinking stays in context (the default on newer Claude models, and on OpenAI GPT-5.6+, which by default renders reasoning from earlier turns), later steps also pay for it again as input.

How to keep reasoning costs under control

1. Measure it. Log reasoning_tokens (OpenAI) or thinking_tokens (Anthropic) for every call. Most surprise bills come from never looking at this field.
2. Lower reasoning effort or budget. Use OpenAI's reasoning.effort, Anthropic's adaptive thinking or budget_tokens (minimum 1,024; Anthropic describes it as a target for reasoning tokens), or Gemini's thinking settings. Save deep thinking for steps that need it.
3. Cap total output. OpenAI's max_output_tokens and Anthropic's max_tokens include reasoning. One catch: OpenAI warns that if you hit the cap during reasoning, the response can come back incomplete before any visible output, and you're still charged for the input and reasoning.
4. Route by difficulty. Let a cheap model (Haiku 5.5, GPT-6 Luna, Gemini 3.8 Flash) handle simple tool-routing steps, and send only the hard planning steps to a frontier model.
5. Manage carried-over thinking. On Claude, thinking block clearing can drop old thinking from context so you don't keep paying for it as input.
6. Batch what isn't urgent. All three providers offer a 50% batch discount, and Anthropic's covers both input and output tokens, so thinking is discounted too. Anthropic also recommends batch processing for thinking budgets above 32K. See Is the Batch API worth it for AI agents?

Estimate it before you ship

To budget an agent honestly, include an assumption for reasoning tokens per step, not just the visible reply. The TokenTab AI agent cost calculator lets you set steps, tokens per step, and model, then compare monthly cost across OpenAI, Anthropic, Google, and others. Add your measured reasoning tokens to the output side and the estimate will match your invoice much more closely.

Sources

Prices and docs retrieved 2026-10-25.

• OpenAI reasoning models guide (billed as output, usage fields, max_output_tokens, incomplete responses, reasoning carry-over on GPT-5.6+): https://developers.openai.com/api/docs/guides/reasoning
• OpenAI API pricing: https://developers.openai.com/api/docs/pricing
• OpenAI Community, reasoning token billing example: https://community.openai.com/t/how-are-reasoning-tokens-cached-tokens-input-tokens-and-output-tokens-counted-for-billing/1386849
• Anthropic thinking docs: https://platform.claude.com/docs/en/build-with-claude/thinking
• Anthropic, steering thinking and cost (billing table, thinking_tokens): https://platform.claude.com/docs/en/build-with-claude/thinking-steering-and-cost
• Anthropic extended thinking (budget_tokens rules, batch above 32K): https://platform.claude.com/docs/en/build-with-claude/extended-thinking
• Anthropic context windows (thinking block preservation by model): https://platform.claude.com/docs/en/build-with-claude/context-windows
• Anthropic pricing (model rates, batch discount on input and output): https://platform.claude.com/docs/en/about-claude/pricing
• Google Gemini API pricing ("Output price (including thinking tokens)", promo dates): https://ai.google.dev/gemini-api/docs/pricing
• TokenTab models.json: /models.json