← Other blogs

Claude Sonnet vs Opus cost: how to choose a default

Double the rate per token, a worked month at both prices, and the cases where the cheaper tier loses.

By ClaudeGPT Si Newsroom · Editorial team

· 3 min read

Claude Sonnet vs Opus cost comes down to one ratio: Opus charges twice as much per token. Sonnet 5.5 lists at $2 per million input tokens and $10 per million output tokens. Opus 5.5 lists at $4 and $20. Those are Anthropic's prices as of November 2026, and they can change.

The rates, side by side

Line item Sonnet 5.5 Opus 5.5
Input, per million tokens $2 $4
Output, per million tokens $10 $20
Cache read, per million tokens $0.10 $0.20
5 minute cache write, per million tokens $2.50 $5

A cache write costs 1.25 times the base input price, which is why it sits above the plain input row on both models.

A month of work, priced both ways

Suppose one developer sends 2 million input tokens and 500 thousand output tokens in a month, with no caching. On Sonnet that is $4 for input and $5 for output, so $9. On Opus it is $8 and $10, so $18. Double the work and both totals double. The ratio does not move, which is what makes the choice easy to model and hard to get wrong on price alone.

Why the bill gap is not always double

The ratio holds for identical token counts. Real work rarely is identical. A task that Sonnet finishes in one pass and Opus finishes in one pass costs twice as much on Opus. A task that Sonnet needs three attempts to finish can cost more on Sonnet than a single Opus run, because each attempt sends the context again. The model that ends the loop sooner can be the cheaper one, and only your own logs show which is happening.

What the rate leaves out

Thinking is part of the output count, so a model that reasons longer bills more even when the answer is the same length. A subscription replaces per-token billing entirely: Pro is $20 a month, or $17 with the annual discount, Max is $100 and Ultra is $200, and all three include Claude Code. On a seat, the question is how much work fits before a limit, not what a token costs. Anthropic also says the model in use changes how fast a limit arrives.

A default that holds

Start on Sonnet. Raise to Opus for a task that has already failed, is expected to run unattended for a long stretch, or touches a large unfamiliar codebase. Drop to Haiku 5.5, from $0.10 per million input tokens, for short mechanical jobs such as renaming or formatting. Then read a week of real usage before changing the policy again.

Checking the choice

UsageAtlas shows tokens by model and estimated cost from local logs, free and open source, with no account needed for local tracking and no reading of prompts or responses. It is a quick way to confirm that the default you set is the model doing the work. Estimates stay estimates, and the Anthropic account is the bill.

Model choice also changes how many turns a task takes, and every turn resends the conversation. That is why two teams with the same default can report very different monthly figures. Write the policy down where the team can see it, because an unwritten default drifts with whoever set up the last session.

One last check before you switch everything: price the task, not the token. A default is a starting point, and the exceptions are where the money is decided.