Claude Sonnet 5.5 Keeps Token Prices but Changes the Cost per Task
Claude Sonnet 5.5 launched on 28 September at the same API rates as Sonnet 5: $2 per million input tokens and $10 per million output tokens. Anthropic says it generates output more than 30% faster and can cost up to 30% less per task because it often uses fewer tokens. Developers should test their own workloads and review the new thinking setting before migrating.
On this page
Sonnet 5.5 keeps the old token rates
Anthropic released Claude Sonnet 5.5 on 28 September with unchanged API rates for its midrange model: $2 per million input tokens, $10 per million output tokens, and $0.20 per million cached input tokens. In its launch announcement, the company says the model generates output more than 30 per cent faster than Sonnet 5 and costs up to 30 per cent less for most work. Those are Anthropic's measurements, not a price cut on the rate card.
The distinction matters to anyone budgeting an agent. A task that ends in fewer tokens, tool calls, and retries can cost less at identical per-token prices. A task that runs at a higher effort setting or spends longer checking its work can cost more. Reuters independently confirmed the unchanged $2 and $10 rates, but the claimed speed and savings still need testing on a buyer's own prompts and tools.
Sonnet 5.5 is the second model in the 5.5 family after Opus 5.5. Opus costs $4 per million input tokens and $20 per million output tokens. Anthropic describes Sonnet as suited to well-scoped coding and everyday document work, while recommending Opus for complex, open-ended work that requires sustained judgment. That is a useful starting hypothesis for routing tasks, not a guarantee that the cheaper model will finish every job for half the cost.
The headline benchmark jump needs context
Anthropic reports a 70.6 per cent score for Sonnet 5.5 on Terminal-Bench 4.0, compared with 10.3 per cent for Sonnet 5. The difference is striking, but it is a vendor-reported result for a particular agent setup. It does not establish the same improvement in every coding repository or prove that a successful patch will survive a project's tests and review.
The company's own release also narrows its competitive claims. Sonnet 5.5 approaches Opus 5.5 on some listed evaluations, yet Anthropic says Opus remains stronger on complex work requiring judgment over a long run. On one coding evaluation, Sonnet 5.5 scored lower at Max effort than at Xhigh because extra review steps led to timeouts or changes beyond the task's scope. More reasoning and more tools are not automatically better.
The practical comparison is a fixed set of representative tasks with the same success criteria. Record completion rate, human corrections, elapsed time, total tokens, and tool calls. Compare the cost of a finished acceptable result rather than the price of a million tokens in isolation. For local-model users, this is also the fair way to compare a hosted model with a device-run workflow: include the work each system actually completes, as well as the different privacy and availability boundaries.
Migration changes and safety fallbacks
Developers can call claude-sonnet-5-5 on Anthropic's platform, and Anthropic says the model is available through AWS, Google Cloud, and Microsoft Azure. The release says developers who previously ran Sonnet with thinking off need to switch to the between_tools setting before migrating. Claude apps and Claude Code default to Medium effort; Anthropic's developer platform defaults to High, so an unchanged prompt may not imply an unchanged work budget across surfaces.
The migration guide adds a cost trap for teams coming from Sonnet 4.6 or earlier. The same text can use roughly 30 per cent more tokens under Sonnet 5.5's newer tokenizer, and higher-resolution image handling can make image-heavy prompts more expensive. Its lower-cost-per-task claim compares against Sonnet 5, not every older Sonnet. Recount tokens and rerun the workload before setting a new budget.
Anthropic says Sonnet 5.5's cybersecurity capabilities are much higher than Sonnet 5's. It therefore applies safeguards similar to those on Opus 5.5, with higher-risk cyber requests visibly falling back to Sonnet 5. Routine software development is intended to continue, but teams using Claude for authorised security testing should verify whether their specific tasks encounter the fallback. The model also adds controls against large-scale extraction of its reasoning.
Sonnet 5.5 is a hosted model. Its availability does not alter the storage or network boundary of a local AI app. For teams choosing between the two, the new release chiefly changes the hosted option's potential speed and cost. Its actual advantage is measurable only after running the same work under the same acceptance checks.