Claude Sonnet 5.5 ships at $2 and $10
On 28 September 2026 Anthropic released Claude Sonnet 5.5, the second model in its Claude 5.5 family. Developers call it as claude-sonnet-5-5 on the Claude Platform. Anthropic says it is available on all platforms, including Amazon Web Services, Google Cloud and Microsoft Azure. Zero data retention is available, as with Sonnet 5.
Anthropic says Sonnet 5.5 generates output more than 30% faster than Sonnet 5. It says the model costs up to 30% less per task because it needs fewer tokens. Anthropic says input, output and cache-read prices match Sonnet 5. Per million tokens, input costs $2, output costs $10 and cache reads cost $0.20. Cache writes cost $2.50. Opus 5.5 costs $4 for input and $20 for output.
Anthropic positions Sonnet 5.5 for well-scoped everyday work, bug fixes and documents. It says Claude Haiku 5.5 will follow "in the coming weeks".
Anthropic and every Sonnet 5 integration
- Anthropic shipped the model, its system card and new safeguards.
- Developers on the Claude Platform, AWS, Google Cloud and Azure can switch model ids now.
- Claude Code and Claude app users get Medium effort by default. The Claude Platform defaults to High effort.
- Cyber defenders can soon apply to an expanded Cyber Verification Program for tiered access on Sonnet 5.5, Opus 5.5 and Claude Mythos models.
70.6% on Terminal-Bench 4.0 against 10.3%
Anthropic published this table. All scores are Anthropic's own reports or third-party runs that Anthropic cites.
| Benchmark | Sonnet 5.5 | Sonnet 5 | Opus 5.5 | GPT-6 Sol |
|---|---|---|---|---|
| Terminal-Bench 4.0 | 70.6% | 10.3% | 66.4% (Xhigh) | not shown |
| FrontierCode 1.1 (Main) | 46.2% (Max), 52.1% (Xhigh) | 42.4% | 54.4% | 49.3% |
| CursorBench 4.0 | 55.5% | 34.1% | 57.8% | not shown |
| GDPval-AA v2.1 | 1844 | 1449 | 1846 | 1487 |
| AA-Briefcase v1.1 | 1811 | 1359 | 1822 | 1483 |
| Humanity's Last Exam, with tools | 64.5% | 54.9% | 67.7% | not shown |
| OSWorld 2.1 (partial) | 80.1% | 57.0% | 81.8% | not shown |
| Chartography, no tools | 61.6% | 15.6% | 64.4% | 53.6% |
Anthropic says Sonnet 5.5 is the first Sonnet to launch with cyber safeguards. It says the model's cyber skills are comparable to Opus 5. Higher-risk cyber tasks now fall back to Sonnet 5, and the user sees the fallback. Anthropic says users can still find and fix bugs as part of routine development. Biology safeguards stay the same as Sonnet 5.
Anthropic also calls it the first Sonnet with safety classifiers that prevent reasoning extraction. Anthropic links this to distillation attacks that use thousands of fake accounts. Preserved thinking now ties Claude's thinking to the account that created it. The docs say the API drops Sonnet 5.5 thinking blocks sent from another, unlinked account. The model then answers without that reasoning.
Anthropic ran an automated behavioral audit of about 1,850 scenarios. It says Sonnet 5.5 improves on or matches Sonnet 5 on most alignment, misuse and honesty measures. On containment tests, Anthropic says it comes close to Opus 5.5. Anthropic says it is the least likely of its models to probe the limits of its containers.
Vendor scores and an Opus 5.5 ceiling
- Anthropic says Opus 5.5 remains clearly stronger at complex, open-ended work that needs sustained judgment.
- The "up to 30%" cost cut comes from Anthropic's own testing. Your token savings depend on your prompts and effort setting.
- Artificial Analysis ran GDPval-AA and AA-Briefcase on a pre-release deployment with a structured-output bug. Anthropic says the bug is fixed and expects any effect to be small.
- Anthropic notes that OpenAI recently fixed an image bug in GPT-6 Sol. Third-party scores for GPT-6 Sol may not reflect that fix yet.
- The post gives no release date or price for Haiku 5.5.
Migrate thinking-off calls to between_tools
- Teams that run Sonnet with thinking off must switch to the new
between_toolssetting before moving to Sonnet 5.5. Follow the migration guide. - Teams that move sessions between accounts, including Claude Code users who switch accounts mid-session, should read the preserved-thinking docs first.
- Security teams whose work touches offensive tooling should test for Sonnet 5 fallbacks. Apply to the Cyber Verification Program if fallbacks block legitimate work.
- Cost owners should rerun one real workload at Medium and High effort. Compare cost per task before and after the swap.