Anthropic released Claude Sonnet 5.5 on September 28, 2026, at the same $2 input and $10 output per million tokens as Claude Sonnet 5. Anthropic says it runs more than 30% faster and costs up to 30% less per task because it uses fewer tokens. On October 7, 2026, Anthropic also halved its cache-read price to $0.10 per million tokens, so the launch-day price table is out of date on that line.
Sonnet 5.5 is the second model in the Claude 5.5 family, after Claude Opus 5.5 and before Claude Haiku 5.5. In its announcement, Anthropic describes it as the faster, lower-cost complement to Opus 5.5: strongest at well-scoped everyday tasks, bug fixes and document, slide and spreadsheet work, with Opus 5.5 kept for complex work that needs sustained judgment.
Anthropic's models overview lists a 1M-token context window, 128K maximum output tokens, a June 2026 knowledge cutoff, a default effort level of high, and retirement no sooner than September 28, 2027. The API identifier is claude-sonnet-5-5, with no date suffix. It is available on the Claude Platform, Amazon Web Services, Google Cloud and Microsoft Azure, and Anthropic says it is offered with zero data retention, as Opus 5.5 and Sonnet 5 are.
| Per million tokens (USD) | Claude Sonnet 5.5 | Claude Opus 5.5 |
|---|---|---|
| Input | $2 | $4 |
| Output | $10 | $20 |
| Cache writes (5-minute) | $2.50 | $5 |
| Cache writes (1-hour) | $4 | $8 |
| Cache reads | $0.10 | $0.20 |
Sources: Anthropic pricing documentation and claude.com/pricing, checked October 10, 2026.
The cache-read cut was announced in the Claude Haiku 5.5 post on October 7, 2026, which says it reduces the cost of Sonnet 5.5 on most agentic tasks by around 20%. The Sonnet 5.5 launch page still shows $0.20, so check the date on any comparison you read.
The practical effect is that Sonnet 5.5 now costs exactly half of Opus 5.5 on every line of the price sheet. For agent loops that re-read a large cached context, that makes the choice between them mainly a question of task success rate, not token price.
Anthropic's announcement compares Sonnet 5.5 with Sonnet 5, Opus 5.5 and GPT-6 Sol. The most striking figure is Terminal-Bench 4.0, an agentic coding evaluation: Anthropic reports 70.6% for Sonnet 5.5 against 10.3% for Sonnet 5, and 66.4% for Opus 5.5 at its highest-scoring effort level. On most other evaluations in the table, including CursorBench, OSWorld and Humanity's Last Exam, Sonnet 5.5 scores a little below Opus 5.5.
Anthropic adds its own caveat: in its testing and that of external testers, Opus 5.5 "remains clearly stronger at complex, open-ended work requiring sustained judgment." These are Anthropic's measurements, and this article does not reproduce them. The table is still useful as a map of where the smaller model is close to the larger one (terminal-style coding, chart reading, computer use) and where Anthropic expects a gap.
The Sonnet 5.5 migration guide lists changes that apply to every earlier Sonnet model.
tool_choice of any or a named tool returns a 400 error. The guide's replacement is tool_choice: auto with strict: true on the tool, plus a prompt instruction saying when to use it. Strict tool use is limited to 20 tools per request, and the guide notes that on Amazon Bedrock structured outputs, including strict tool use, are not available for Sonnet 5.5, so code there must validate tool inputs itself.computer_20251124 returns a 400 error there; use computer_toolset_20260801.There is also a behavior change for teams that ran Sonnet without thinking. On Sonnet 5.5, a request with no thinking field runs with adaptive thinking, and disabled is no longer accepted. The lowest setting is thinking: {"type": "between_tools"}, which turns off up-front thinking but still returns progress updates between tool calls as thinking blocks. Code that reads content[0].text will break when a response starts with a thinking block, and thinking tokens count against max_tokens and are billed as output.
between_tools setting and a fresh latency test.Teams on Sonnet 5 pay the same input and output prices for a faster model with cheaper cache reads. Plan for the forced-tool-use and thinking changes.
Teams defaulting to Opus for everything can test whether well-scoped tasks hold up on Sonnet 5.5 at half the price. Anthropic itself suggests keeping Opus 5.5 for open-ended, judgment-heavy work.
Teams comparing vendors will find that OpenAI's GPT-6.1 Sol also lists at $2 input and $10 output per million tokens, with $0.10 cached input, according to its model page. With identical headline prices, the comparison comes down to cache rules (OpenAI's prompt caching update gives discounts for shared prefixes reused within a 30-minute window, against Anthropic's 5-minute and 1-hour cache writes), long-context pricing (the same OpenAI model page bills requests above 272,000 input tokens at higher rates for the whole request), and tokens used per task. See the GPT-6.1 Sol guide for detail.
claude-sonnet-5-5.tool_choice with auto plus strict tools, or validate inputs in code on Bedrock.thinking: disabled with between_tools if you need the lowest thinking setting.max_tokens to cover thinking.computer_toolset_20260801 on the Claude API and Google Cloud.When was Claude Sonnet 5.5 released?
September 28, 2026. Its cache-read price was cut on October 7, 2026.
How much does Claude Sonnet 5.5 cost?
$2 input, $10 output, $2.50 for 5-minute cache writes, $4 for 1-hour cache writes and $0.10 for cache reads per million tokens, per Anthropic's pricing documentation (checked October 10, 2026).
Is Claude Sonnet 5.5 better than Claude Opus 5.5?
Anthropic reports Sonnet 5.5 ahead on Terminal-Bench 4.0 and slightly behind on most other evaluations, and says Opus 5.5 is stronger for complex, open-ended work. Sonnet 5.5 costs half as much.
Discover more content: