What Anthropic Announced
Anthropic announced Claude Sonnet 5.5 on September 28, 2026, as the second model in the Claude 5.5 family, following Opus 5.5. The company positions it for well-scoped everyday tasks, such as fixing code bugs and preparing documents, presentations, and spreadsheets. Haiku 5.5, designed for high-volume and cost-sensitive applications, was announced for release in the coming weeks.
The company describes Sonnet 5.5 as a complement to Opus 5.5, not a universal replacement. The intended use is to deploy Sonnet for frequent, bounded assignments and reserve Opus for open-ended tasks requiring sustained judgment. This recommendation reflects Anthropic's positioning; the practical choice depends on the specific workflow, required quality, and total cost.
Speed and Cost: Not the Same as a Lower Rate
Anthropic states that Sonnet 5.5 generates responses more than 30% faster than Sonnet 5 and can cost up to 30% less per task in its testing. However, the published price per million input and output tokens is the same as for Sonnet 5: $2 for input and $10 for output. According to the company's explanation, the per-task savings come from the model typically needing fewer tokens to complete the work, not from a general reduction in those unit rates.
Effective spending varies depending on context length, response size, tool calls, and the chosen effort level. A faster result does not automatically mean a complete task takes 30% less time: there may be waits, reviews, and external steps. To assess the savings, one must measure the cost and time of completed tasks using comparable criteria.
Pricing published by Anthropic
| Item | Claude Sonnet 5.5 | Claude Opus 5.5 |
|---|---|---|
| Input, per million tokens | $2 | $4 |
| Output, per million tokens | $10 | $20 |
| Cache read, rate shown in announcement | $0.20 | $0.20 |
| Cache write, rate shown in announcement | $2.50 | $5 |
What the Benchmarks Say and What They Don't Prove
Anthropic reports a 70.6% score for Sonnet 5.5 on Terminal-Bench 4.0, compared to 10.3% for Sonnet 5 and 66.4% for Opus 5.5. On CursorBench 4.0, it reports 55.5% for Sonnet 5.5, 34.1% for Sonnet 5, and 57.8% for Opus 5.5. These figures allow describing the results the company attributes to its models in those tests; by themselves, they do not establish how the models will respond to a specific task or prove that the difference replicates in other environments.
Comparisons depend on configuration. Anthropic publishes results at different effort levels and warns that scores can change based on that setting. It also reports that Sonnet 5.5 scored lower on FrontierCode at Max effort than at Xhigh; it attributes this partly to additional reviews that in some cases caused timeouts or out-of-scope changes. For GDPval-AA and AA-Briefcase, the company notes that a pre-release version with a bug related to structured outputs was evaluated, which was later fixed.
No independent, reproducible tests are provided that allow confirming the speed and per-task savings claims under general conditions. A useful evaluation should specify version, effort, tools, task, repetitions, and quality criteria; it must also account for total time, tokens, tool calls, and human review.
A Practical Process for Testing
An independent comparison can clarify whether the assumed savings are worth it for a specific application. The test set should represent real tasks, include simple and difficult cases, and keep the same instructions, data, and tools for each model. If possible, repeat each case: a single run can vary and is not enough to estimate typical behavior.
Evaluate the complete outcome, not just the first response. Record whether the task was resolved correctly, how long it took from start to finish, how many tokens it consumed, what tool calls it made, and how much time a person needed to review or correct it. A lower per-task rate does not imply an improvement if it increases rework or the risk of errors.
Comparable test before switching models
- 01Select representative tasks and define in advance what counts as a correct result.
- 02Run Sonnet 5 and Sonnet 5.5 with the same instructions, tools, and data; note the version and effort level.
- 03Repeat the cases to detect variations and record total time, tokens, tool calls, and errors.
- 04Have a person review the results without knowing, if possible, which model generated them.
- 05Compare cost per accepted task and correction work, and test Opus 5.5 when the assignment is open-ended or requires sustained judgment.
Availability, Access, and Safeguards
Anthropic indicates that Sonnet 5.5 is available on its platforms, including Claude, Amazon Web Services, Google Cloud, and Microsoft Azure. The announced identifier for the Claude Platform is 'claude-sonnet-5-5'. AWS confirms availability on Amazon Bedrock and Claude Platform on AWS in its publication. The provided information does not allow specifying whether access is offered in the same way across all plans, regions, or accounts.
Anthropic also states that the model is available with zero data retention and incorporates cybersecurity safeguards similar to those of Opus 5.5. According to the company, certain higher-risk requests may trigger a fallback to Sonnet 5, while routine software development should not be affected. The firm claims that biology safeguards remain the same as in Sonnet 5 and warns that some requests might be flagged in error. These are provider descriptions, not an independent assessment of their effectiveness.
What Remains to Be Verified
The speed and per-task cost figures are Anthropic's claims based on its evaluations. The provided material does not include independent results with full methodology that allow establishing how well they replicate in other products or workloads. There is also a lack of current, applicable primary documentation to resolve the post-announcement cache read rate.
The prudent conclusion is that Sonnet 5.5 presents improvements published by its manufacturer and is available on the indicated platforms, but production performance must be verified with your own tasks. When comparing, it is not advisable to mix different configurations, versions, or markets, nor to take an isolated benchmark as definitive proof. The choice between Sonnet and Opus will depend on the observed balance between quality, cost, speed, and review.
Initial decision based on work type
| Situation | What to test first | What to measure |
|---|---|---|
| Frequent, well-scoped task | Sonnet 5.5 | Correction, total time, and cost per accepted task |
| Work with multiple open decisions | Compare Sonnet 5.5 and Opus 5.5 | Result quality, errors, and necessary review |
| Automated flow with tool calls | Sonnet 5.5 in a controlled test | Tokens, calls, failures, and time to completion |
| Migration of an existing application | Test in a separate environment before switching | Configuration compatibility and regression results |
Open questions
- No independent, reproducible tests are provided that confirm the speed or per-task cost claims under general workloads.
- The available material does not resolve the post-announcement cache read rate with current primary documentation: the launch page shows $0.20 per million tokens, while another reference claims it dropped to $0.10 starting October 7, 2026.
- Availability may vary by provider, plan, account, or region; the provided sources do not detail all these conditions.
- Benchmarks depend on versions, effort levels, and evaluation methods; they should not be treated as a guarantee of production results.
Keep exploring
Sources consulted
Corrections and transparency
If you spot incorrect or outdated information, send us a correction with the page and source we should review.
Submit a correction