The same rates, a claimed smaller bill

Anthropic released Claude Sonnet 5.5 on September 28, positioning it for everyday coding and document work. The company says it generates output more than 30% faster than Sonnet 5 and can cost up to 30% less per task in its tests. Neither figure is a guaranteed saving for a particular application.

The distinction is in the bill. Anthropic's pricing page lists the same standard rates for Sonnet 5 and 5.5: $2 per million input tokens and $10 per million output tokens. A token is a small piece of text processed by a model. Cache reads, which reuse previously stored input, cost $0.20 per million tokens for both models.

Anthropic says the new model needs fewer tokens to finish work. That can reduce a bill without changing the price of each token. It is not a blanket discount, and API rates should not be confused with the price of a Claude subscription.

A name change is not the whole upgrade

The release is accompanied by a fairly consequential migration guide. Developers using the Messages API must do more than replace the model identifier with claude-sonnet-5-5 if their existing requests rely on settings the new model rejects.

One example is thinking, the model's additional reasoning before or between actions. The disabled setting returns an error on Sonnet 5.5. To keep up-front thinking off, Anthropic directs developers to between_tools at low, medium or high effort. That option can still produce reasoning updates between tool calls; it does not mean every kind of thinking has been removed.

Forced tool choice is another change. Requests that insist on a particular tool, or any tool, are rejected. The guide instead describes automatic tool selection and, where supported, strict validation of tool inputs. Automatic selection does not itself guarantee a tool will be called. These details matter for software that expects a fixed sequence of actions.

Measure the finished task, not just the response

The company's published comparisons are useful starting points, not proof that every workflow improves. Model Current has not independently run the model or reproduced the speed and cost results.

Our practical reading is to compare the same real tasks with the same acceptance criteria. Record whether the result works, how often it needs another attempt, how long completion takes and what the entire run costs. A shorter answer is not a saving if somebody has to repair it.

Anthropic's pricing documentation also notes that server-side tools can carry additional usage charges. That is another reason to look at the whole task rather than multiply the length of one reply by a headline token rate. The open question is how the new model behaves in a reader's own workload, including failures and review time.

Sources

  1. Anthropic: introducing Claude Sonnet 5.5September 28, 2026 release. Speed and per-task cost reductions are Anthropic's claims, not independently reproduced results.
  2. Claude Platform: published API pricingCurrent pricing table confirms identical Sonnet 5 and 5.5 standard input, output and cache-read rates, plus separate server-side tool charges.
  3. Claude Platform: Sonnet 5.5 migration guidePrimary technical documentation for model ID, between_tools effort limits, disabled-thinking rejection and forced tool-choice changes.