Claude Sonnet 5.5 is designed to make common coding and knowledge-work tasks faster and cheaper to run. Anthropic says it generates output more than 30% faster than Sonnet 5 and costs up to 30% less per task in its testing. For API customers, the distinction is that Anthropic has not cut its per-token prices.
Announced on September 28, Sonnet 5.5 sits in the middle of Anthropic’s Claude 5.5 lineup. It is the company’s faster, lower-cost option for well-scoped work; Opus 5.5 remains its choice for difficult, open-ended problems. The model is available now, though developers moving an existing integration should inspect their thinking, tool-use and response-handling code before changing the model ID.
Sonnet 5.5 Is Built for Repeated Work, Not Every Hard Problem
Anthropic positions Claude Sonnet 5.5 as an upgrade for everyday tasks such as fixing bugs and producing documents, slides and spreadsheets. Its advantage is clearest when an application runs many tasks that can be defined and checked without prolonged judgment.
Opus 5.5 has a different role. Anthropic says its higher-tier model remains clearly stronger on complex, open-ended work requiring sustained judgment, even though Sonnet 5.5 approaches it on some evaluations. A coding assistant handling routine fixes may favor Sonnet’s speed, while an agent making architectural decisions over a long project may still justify Opus.
The choice also depends on how much work the model does. Anthropic’s performance-and-cost charts vary the effort setting, which controls how deeply the model works on a request. At higher settings, Sonnet 5.5 can approach Opus 5.5’s performance, but its cost per task can approach Opus’s too. A team looking for a cheaper default should compare the models at the effort levels it would actually deploy.
Lower Task Cost Does Not Mean Cheaper Tokens
The API list price remains $2 per million input tokens and $10 per million output tokens, matching Sonnet 5. Anthropic also lists cache reads at $0.20 per million tokens. The company attributes its claimed task-level savings to Sonnet 5.5 typically using fewer tokens to complete the same work, not to a new rate card.
Consider a request containing 100,000 normally billed input tokens and producing 10,000 output tokens. It would cost $0.30 at the listed rates, before any applicable caching or other charges. If an equally successful run kept the input fixed but reduced output to 7,000 tokens, its token charge would be $0.27. Saving 30% of the output tokens saves 10% of the total token bill in this example because input remains part of the cost.
Anthropic’s “up to 30% less per task” figure comes from its tests, not a discount on every request. An agent may save money by finishing in fewer steps, issuing fewer tool calls or generating less output. Another workload may see little change. The company says early testers observed more batched tool calls than with Sonnet 5, but developers need to measure completed tasks in their own applications to see whether that translates into savings.
The speed claim also has a boundary. Anthropic says Sonnet 5.5 generates outputs more than 30% faster than Sonnet 5. That helps with interactive work, but an agent also spends time waiting for tools, external services and repeated model calls. Faster generation does not promise that every end-to-end workflow will finish 30% sooner.
Anthropic’s Benchmarks Show Gains With Important Boundaries
Anthropic reports substantial improvements over Sonnet 5 in coding and knowledge-work evaluations. On its published CursorBench 4.0 results, which assess ambiguous, multi-file coding tasks drawn from Cursor sessions, Sonnet 5.5 scores 55.5%, compared with 34.1% for Sonnet 5 and 57.8% for Opus 5.5. On Terminal-Bench 4.0, an agentic command-line evaluation, Anthropic reports 70.6% for Sonnet 5.5 against 10.3% for Sonnet 5.
The reported gap with Opus is narrower on knowledge work. Anthropic lists Sonnet 5.5 at 1,844 on GDPval-AA v2.1, against 1,846 for Opus 5.5 and 1,449 for Sonnet 5. Its account of that benchmark describes tasks across occupations and industries, but a close aggregate score does not establish that the two models make equally good decisions on every open-ended assignment.
These are Anthropic-reported evaluations, not independent confirmation of what a customer will experience. Agent setup, effort setting and the cost of repeated attempts affect the practical comparison. The striking Terminal-Bench result is worth investigating, but it does not guarantee the same jump on an organization’s own codebase.
A developer deciding whether to switch could hold the task set and success criteria steady, then record completion rate, time and total billed tokens at several effort levels. That would show whether Sonnet 5.5 reaches an acceptable result more efficiently than Sonnet 5 or Opus 5.5, beyond what a benchmark score alone can tell them.
Changing the Model ID Is Only the First Migration Step
The Claude API model ID is claude-sonnet-5-5, without a date suffix. Anthropic says Sonnet 5.5 is available through its platform and on AWS, Google Cloud and Azure. Developers using another provider should check that provider’s model identifier and supported features rather than assuming the Claude API string applies unchanged.
For code already running Sonnet 5, Anthropic’s migration guide identifies several changes that can produce errors or alter an agent’s behavior:
- Review thinking settings. Adaptive thinking is on by default, with high effort as the Claude API default. Sonnet 5’s
thinking: {"type": "disabled"}returns a 400 error on Sonnet 5.5. The lowest setting isthinking: {"type": "between_tools"}, available at low, medium and high effort; it does not eliminate thinking between tool calls. - Remove forced tool choice. Sonnet 5.5 rejects
tool_choicevalues that require any tool or a named tool. Use automatic tool choice and, where supported, strict tool schemas for valid inputs. That does not guarantee the model will call a tool, so applications must still handle a text response. Anthropic says strict tool use is unavailable for this model on Amazon Bedrock. - Check computer-use integrations. On the Claude API and Google Cloud, Sonnet 5.5 rejects the older
computer_20251124tool and requirescomputer_toolset_20260801. Amazon Bedrock still accepts the older tool, according to Anthropic’s documentation. - Handle thinking blocks in responses and history. Code that assumes the first response block contains text can fail. Tool loops must pass thinking blocks back unchanged. Anthropic also says text written between tool calls can arrive in thinking blocks, changing what an application displays to users.
- Recheck long-running conversations. Thinking blocks are tied to the model and conversation that produced them. Sonnet 5.5 can read Sonnet 5’s thinking blocks, but switching away from Sonnet 5.5 does not carry its reasoning blocks into another model. On some accounts, editing earlier conversation content before replaying a Sonnet 5.5 thinking block can also trigger an error.
Frequently Asked Questions
4 questions
1Is Claude Sonnet 5.5 Cheaper Than Sonnet 5?
Claude Sonnet 5.5 has the same listed API rates as Sonnet 5: $2 per million input tokens and $10 per million output tokens. Anthropic says Sonnet 5.5 costs up to 30% less per task in its tests because it typically uses fewer tokens to finish the work. Actual savings depend on the task, effort setting and tokens billed.
2
Sources
- Claude Sonnet 5.5anthropic.com
- Anthropic’s migration guideplatform.claude.com
- Sonnet 5.5 feature documentationplatform.claude.com
