Claude Sonnet 5.5: maximum effort can cost more than Opus

Claude Sonnet 5.5 launched on 28 September 2026 with faster output and the same token prices as Sonnet 5. Independent results show why I would still choose its reasoning effort carefully: Artificial Analysis measures a higher cost per task for Sonnet 5.5 at maximum effort than for Opus 5.5 at maximum effort. The cheaper token is not always the cheaper task. [1] [2] [7] [9] [10]

Did Anthropic cut Sonnet 5.5’s prices?

The token prices are unchanged. Anthropic lists $2 per million input tokens and $10 per million output tokens, with cache reads at $0.20. [9] Its claim of up to 30% lower task cost comes from using fewer tokens in its own tests, rather than charging 30% less for each token. [1]

That is an important difference for an agent that decides how long to work. A request can get cheaper at unchanged rates if the model reaches an acceptable result with fewer tokens. It can also get more expensive if you ask it to spend longer checking, trying alternatives or delegating work. The provider’s price table tells you what a token costs; it does not tell you how many you will buy.

The release also advertises at least 30% faster output generation. That measures how quickly text arrives, rather than how quickly a whole job finishes. A repository task can spend time waiting for tests, calling tools or revising a failed change, so faster generation is a useful improvement without being an end-to-end timing result. [1]

I would compare this release with Opus 5.5’s measured capability and cost using completed work. The token discount relative to Opus matters, but only after the model has produced something I can accept.

What changes when you raise Sonnet’s effort?

The independent results show a substantial trade-off between aggregate score and measured spending. Artificial Analysis evaluates Sonnet 5.5 with adaptive reasoning and default fallback at four effort settings. Its current Intelligence Index v4.3.2 rises from a rounded 41 at medium to 56 at maximum, while the measured cost per benchmark task rises from $0.59 to $7.60. [2] [3] [4] [5]

Maximum effort gets the highest aggregate score in this comparison, but costs about thirteen times as much per task as medium. That is my calculation from the evaluator’s $7.60 and $0.59 figures. It gives me a reason to increase effort when a harder task needs it, rather than use the highest setting for every change.

Artificial Analysis’ cost is a weighted benchmark measurement using the tokens consumed and typical cache-hit rates. The index combines ten mostly English, text-based evaluations. These results compare settings within that suite; they do not estimate a typical Claude subscription user’s bill. [6]

The launch has a further evaluation caveat. Anthropic says Artificial Analysis ran GDPval-AA and AA-Briefcase on a prerelease deployment with a structured-output bug that was subsequently fixed. The company expects any effect to be small and to understate performance. I would keep the retrieved figures dated rather than present them as a permanent ranking. [1]

How can Sonnet cost more than Opus?

Sonnet 5.5 at maximum effort costs $7.60 per Artificial Analysis task, compared with $5.98 for Opus 5.5 at maximum effort. Sonnet’s lower token rates therefore do not translate into a lower task cost in this particular comparison. It is the amount of model work, as well as the rate, that determines the result. [2] [7]

That should change how you read the word “cheap” in a model announcement. It might mean inexpensive input, inexpensive output, or a low cost to finish a specific job. Those are related measurements, but they answer different questions. The effort table makes the distinction visible without requiring us to guess how a model will behave in an unfamiliar repository.

This is also why benchmark scores alone miss some of model quality. A higher score may be useful enough to justify more spending on a difficult task, while adding little value to a straightforward fix. The comparison only becomes a decision when the task and the acceptance criteria are clear.

Which effort would I start with?

I would start ordinary Claude Code work at the documented Medium default, then raise effort when the task needs more analysis. Anthropic’s Claude Academy explains that effort controls the resources spent on a problem, rather than the model’s underlying knowledge. Higher settings can mean extra tokens and waiting, and effort labels are not identical amounts of computation across models. [1] [8]

I have not run that controlled comparison for Sonnet 5.5 yet. I would keep planning and implementation as separate jobs so the expensive reasoning goes to the part that needs it. The setting I would keep is the one that finishes the requested work with the least correction.