# Claude Opus vs Sonnet tokens: where the difference shows up

> Opus costs double per token, while the number of tokens a task takes is set by effort, context and retries.

URL: https://claudegpt.si/news/claude-opus-vs-sonnet-tokens  
Published: 2026-11-18T16:24:48.915Z  
Author: ClaudeGPT Si Newsroom  
Section: AI Tools  
Topics: Claude Code, Anthropic, AI coding costs  
Publisher: ClaudeGPT Si

## Key takeaways

- Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, double the $2 and $10 of Sonnet 5.5 as of November 2026.
- The count per task depends on effort, context size, cache reuse and the number of attempts, so the tier alone does not decide token volume.
- Anthropic positions Sonnet 5.5 for everyday coding and Opus 5.5 for long autonomous runs and large refactors.

Claude Opus vs Sonnet tokens is not a straight comparison. Opus charges double per token, and a call on either model can produce a very different count depending on effort, context and how many attempts it takes. Two things are fixed, and the rest moves with the work.

## The fixed part

Anthropic's prices, read in November 2026:

| Measure | Claude Opus 5.5 | Claude Sonnet 5.5 |
| --- | --- | --- |
| Input, per million tokens | $4 | $2 |
| Output, per million tokens | $20 | $10 |
| Cache read, per million tokens | $0.20 | $0.10 |
| Default effort on the API | medium | high |

Every price line is double on Opus. Effort is a setting rather than a price, and the two models ship with different defaults. Prices and defaults change, so check the current pages before planning around them.

## The moving part

Output includes thinking, and effort decides how much thinking happens. Context is sent again on each call, so a long session multiplies input. Cache reads then lower the price of that repetition without removing it from the count. A model that answers in one pass can log fewer tokens than a cheaper one that needs several rounds.

## When Sonnet is enough

Anthropic describes Sonnet 5.5 as its balance of speed and intelligence, and lists everyday coding, data analysis and agentic tool use among its jobs. Everyday edits, test fixes, small refactors and documentation usually land here. When the spec is clear and the test loop is fast, the cheaper tier finishes the task.

## When Opus earns the price

Opus 5.5 is positioned for multihour autonomous coding, large refactors and vision-heavy work. In those jobs a failed attempt is expensive, and a second Sonnet run can cost more than the first Opus run. Your own log shows whether that is happening, but only if you count attempts per task rather than tokens per day.

## A test you can run in a week

Choose five tasks you repeat weekly. Run them on Sonnet and record attempts and output tokens. Move the two that needed retries to Opus and compare the same measures. Most teams end with a mix: Sonnet carries the volume, and Opus carries the few hard calls where a second try would cost more than the upgrade.

## What changes on a subscription

A seat changes the arithmetic. Pro, Max and Ultra are monthly fees, so a token has no price in the invoice; what it has is a share of the five-hour and weekly allowance. The same per-model view shows which model was running when a window emptied, and that is the question a subscriber actually needs answered.

## Where to read the result

[UsageAtlas](https://usageatlas.com) keeps per-model tokens and estimated cost from local session files, free and open source, with no account needed for local tracking. It does not read prompts or responses, and it shows supported limits and reset times beside the token history. The cost column is an estimate, and the useful part of the per-model view is the comparison rather than the exact dollar.

If you cannot measure attempts, price per token will point you the wrong way. A cheaper model that takes three tries is not cheaper, and a single Opus pass is not automatically the expensive choice.

## Frequently asked questions

### Does Opus use more tokens than Sonnet for the same prompt?

Not necessarily. Both models accept the same input, and output length depends on effort and how much thinking the task triggers. Opus always costs more per token, but a short Opus answer can use fewer tokens than a long Sonnet session.

### Is Sonnet good enough for everyday coding work?

Anthropic positions Sonnet 5.5 for everyday coding, data analysis and agentic tool use, and reserves Opus 5.5 for long autonomous runs and large refactors. Test the tasks you repeat before deciding that one tier covers everything.

### Do cache reads change the comparison?

They are recorded in token totals and priced at 5 percent of base input on both models, so a session with heavy reuse can show a large count on either tier. Compare output tokens and attempts as well as the total.

## Sources

- [Pricing](https://platform.claude.com/docs/en/about-claude/pricing), Anthropic
- [Choosing the right model](https://platform.claude.com/docs/en/about-claude/models/choosing-a-model), Anthropic
- [Claude token usage by model](https://usageatlas.com/guides/claude-token-usage-by-model), UsageAtlas
