Anthropic introduces Claude Sonnet 3.5, 30% faster at the same price
Anthropic has introduced Claude Sonnet 5.5 with improved capabilities for programming, data analysis, and executing multi-step tasks.
Speed and token efficiency
Compared to Claude Sonnet 3.5, the new model generates responses more than 30% faster, and according to Anthropic’s tests, task execution costs can be reduced by up to 30% without changing pricing API.
Claude Sonnet 3.5 integrates tool calls more efficiently and completes tasks in fewer steps, using fewer tokens to achieve the same result.
A million tokens of context
Sonnet 3.5 retains a 1 million token context window and a maximum output length of 128,000 tokens. This allows for processing large documents, codebases, and massive datasets within a single request.
The model accepts text and images, including screenshots, charts, and interfaces.
Improvements in programming
Anthropic recorded the most significant gain in the Terminal-Bench 4.0 test, which evaluates an AI’s ability to work independently with a terminal and perform multi-step tasks. Claude Sonnet 5.5 scored 70.6%, compared to 10.3% for the previous Sonnet 5.
The model can analyze code, fix errors, run checks, and continue working without the need to manually specify every subsequent action.
Working with documents and graphics
Anthropic has also improved the creation of tables, presentations, JSON, and structured documents. According to preliminary testing, Sonnet 3.5 follows complex templates more accurately and creates materials that require fewer manual corrections.
Visual information recognition has improved significantly. In the Chartography test, the model scored 61.6% compared to 15.6% for Sonnet 5. The benchmark evaluates an AI’s ability to understand charts and diagrams without additional tools.
Benchmark results
Anthropic has published comparative results for Claude Sonnet 3.5 and the previous Sonnet 3 across seven benchmarks covering programming, computer use, image analysis, and complex problem-solving.
| Benchmark | Sonnet 5 | Sonnet 5.5 |
|---|---|---|
| Terminal-Bench 4.0 (programming) | 10.3% | 70.6% |
| CursorBench 4.0 (programming) | 34.1% | 55.5% |
| GDPval-AA v2.1 (work tasks, index) | 1449 | 1844 |
| AA-Briefcase v1.1 (work tasks, index) | 1359 | 1811 |
| Humanity’s Last Exam (with tools) | 54.9% | 64.5% |
| OSWorld 2.1 (computer control, partial test) | 57.0% | 80.1% |
| Chartography (chart analysis without tools) | 15.6% | 61.6% |
All metrics are based on results published by Anthropic. GDPval-AA and AA-Briefcase use index scores, while other tests use percentages. A higher result indicates better performance on the corresponding benchmark.
Availability and price
Claude Sonnet 5.5 was released on September 28, 2026, and is now available in the Claude apps, via the official API, Amazon Web Services, Google Cloud, and Microsoft Azure.
Cost of usage via API per 1 million tokens:
- Input tokens – $2;
- Output tokens – $10;
- Reading from cache – $0,20;
- Cache write for 5 minutes – $2,50;
- Cache write for 1 hour – $4.
For developers, the model is available under the identifier claude-sonnet-5-5.
