# Reasoning models: what 'thinking' costs and what it buys

- **Published:** 2026-04-08
- **Updated:** 2026-10-12
- **Category:** Economics & Pricing
- **Author:** Sinthetix
- **URL:** https://sinthetix.com/blog/reasoning-models-what-thinking-costs-and-buys

> Reasoning models spend extra tokens working before they answer. A plain-English look at what that buys, how it is billed, and where it is not worth it.

A standard language model produces its answer in one pass. A reasoning model first spends additional computation working through the problem, then answers. Most current frontier models offer some form of this, and it changes both the quality you can expect on hard tasks and the bill you receive.

**Key takeaways**

- Reasoning models generate extra internal "thinking" tokens before the visible answer.
- Those tokens are billed as output tokens even when you never see them.
- Extra thinking helps most on multi-step problems such as math, coding and analysis, and adds cost and delay on simple requests.
- The practical control is a thinking budget or effort setting, plus measuring tokens used per task.

## What "thinking" actually is

When Anthropic introduced extended thinking in February 2025, it described the model as able to "think longer before answering" and said the mode improves performance on math, physics, instruction-following, coding and other tasks [1]. The thinking is a sequence of intermediate text the model writes while working, before it commits to a final response.

The technique grew out of training. The DeepSeek-R1 paper reported that reasoning behaviors such as self-reflection and verification can be incentivized through reinforcement learning, without human-written step-by-step examples, and that the effect is strongest on verifiable tasks in math, coding competitions and STEM [3]. That also explains where reasoning helps: problems that have a checkable right answer and require several steps.

## How you pay for it

Thinking is not free, and it is not always visible. OpenAI's documentation says reasoning tokens are not returned through the API but still occupy space in the context window and are billed as output tokens; the usage report separates them out [2]. Anthropic's pricing works the same way: its launch post stated the same per-token prices in standard and extended thinking modes, with the output price including thinking tokens [1].

The consequence is that cost depends on how long the model thinks, not just on how long the answer is. The arithmetic below uses Anthropic's published output price of $15 per million tokens [1]. The token counts are illustrative, not measured.

| Scenario (illustrative) | Visible answer | Thinking | Billed output tokens | Output cost |
| --- | --- | --- | --- | --- |
| No extended thinking | 500 | 0 | 500 | $0.0075 |
| Moderate thinking | 500 | 4,000 | 4,500 | $0.0675 |
| Heavy thinking | 500 | 16,000 | 16,500 | $0.2475 |

Per request the amounts are small. At tens of thousands of requests a day, a 9x to 33x multiplier on output is a budget line.

Two practical details from the vendors' documentation matter for planning:

- **Reserve room.** OpenAI recommends reserving at least 25,000 tokens for reasoning and output when you start experimenting, and notes that a request can hit its output cap before any visible answer is produced [2].
- **Set a budget.** Anthropic's API lets you cap thinking with a token budget (minimum 1,024). The budget is a target rather than a strict ceiling, and thinking tokens count toward the request's maximum output. For simple tasks the documentation suggests starting near the minimum; for complex ones, 16,000 or more, with diminishing returns and added latency [4].

## What it buys

Anthropic's launch data illustrates the trade, though it is a vendor result on a specific benchmark. On a 489-task subset of SWE-bench Verified, Claude 3.7 Sonnet scored 63.7% without a high-compute scaffold and 70.3% with one that used parallel attempts, regression-test filtering and ranking [1]. That scaffold is not the same thing as a single request's thinking budget, so read it as evidence that spending more compute per problem can raise accuracy, not as a price list.

Anthropic's documentation makes a related point: larger thinking budgets can improve quality on complex problems, with diminishing returns and added latency [4].

## Where it is worth using

**Good fits:** analyzing a contract against a policy checklist, debugging, reconciling figures across documents, planning multi-step workflows, anything where a wrong answer is expensive to catch later.

**Poor fits:** routing a support ticket, extracting a date from an email, rewriting a sentence, answering from a known FAQ. These tasks rarely improve with extended thinking, and you pay in tokens and waiting time.

A workable policy is to default to the cheapest setting that passes your own test cases, and escalate only the request types where measured accuracy improves.

## Three things to do

1. **Measure thinking tokens per task.** Both vendors report them separately in the usage data [2][4].
2. **Test on your own examples.** A benchmark gain does not tell you whether your invoices or contracts improve.
3. **Cap and monitor.** Set a thinking budget or effort level per workflow, and alert on outliers.

For businesses, treat reasoning depth as a setting you choose per task, not a default you leave on everywhere. The same model can be cheap on simple work and expensive on hard work, and only your own measurements show which tasks are worth the difference.

## Sources

1. [Claude 3.7 Sonnet and Claude Code](https://www.anthropic.com/news/claude-3-7-sonnet), Anthropic, 2025-02-24
2. [Reasoning models](https://developers.openai.com/api/docs/guides/reasoning), OpenAI API documentation, 2026-04-08
3. [DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning](https://arxiv.org/abs/2501.12948), arXiv, 2025-01-22
4. [Extended thinking](https://platform.claude.com/docs/en/build-with-claude/extended-thinking), Anthropic API documentation, 2026-04-08

## How Sinthetix helps

An engineering studio for AI agents, automation, and internal tooling. We build the systems that make operators scale.

[Our services](https://sinthetix.com/services) | [Talk to us](https://sinthetix.com/contact)

Sinthetix is an AI engineering studio. We ship agents, automation, and internal tools for operators who need results, not reports. [About Sinthetix](https://sinthetix.com/about)