One week after Opus 5.5, Anthropic is launching its mid-range model. The Claude Sonnet 5.5 is said to be more than 30 percent faster and cost up to 30 percent less per task. In some tests, it even outperforms its bigger brother.
Anthropic unveiled the Claude Sonnet 5.5 on September 28th. Following the Claude Opus 5.5 from September 22nd, it is the second model in the 5.5 series and replaces the Claude Sonnet 5, which was introduced in June. The key specifications are available in Anthropic's announcement for the Sonnet 5.5.
Sonnet is the middle of the three main lines, positioned between the small Haiku and the large Opus. Above it are the Fable and Mythos models. Haiku 5.5 is expected to follow in the coming weeks.
Key Facts at a Glance
- According to Anthropic, Sonnet 5.5 generates answers more than 30 percent faster than Sonnet 5 and is the fastest Sonnet model to date.
- The price per token remains the same because the model requires fewer tokens, reducing the cost per task by up to 30 percent.
- In several tests, Sonnet 5.5 comes close to Opus 5.5, and in one programming test it even outperforms it.
- As the first Sonnet model, it launches with the same cybersecurity protection mechanisms as the largest models.
Stronger at everyday tasks and programming
Anthropic positions Sonnet 5.5 as a faster and more affordable complement to Opus 5.5. The company sees its strength in clearly defined everyday tasks: fixing code errors, creating documents, presentations, and spreadsheets. It also offers an eye for design, such as revising user interfaces or populating slide templates.
For complex, open-ended work that requires longer deliberation, the Opus 5.5 remains clearly ahead, according to Anthropic. Nevertheless, the test results show how close the mid-range model has come.
| Test | Sonnet 5.5 | Sonnet 5 | Opus 5.5 |
|---|---|---|---|
| Terminal-Bench 4.0 (Programming in the Terminal) | 70,6 % | 10,3 % | 66,4 % |
| CursorBench 4.0 (Programming) | 55,5 % | 34,1 % | 57,8 % |
| GDPval-AA v2.1 (Vocational tasks, points) | 1844 | 1449 | 1846 |
| OSWorld 2.1 (Computer Control) | 80,1 % | 57,0 % | 81,8 % |
| Humanity's Last Exam (with tools) | 64,5 % | 54,9 % | 67,7 % |
All values are taken from Anthropic's announcement. For the terminal benchmark, Anthropic specifies the value for Opus 5.5 at the highest effort level.
Same prices, less consumption
The price per million tokens remains unchanged compared to Sonnet 5. Opus 5.5 costs exactly twice as much per token.
| Price per million tokens | Sonnet 5.5 | Opus 5.5 |
|---|---|---|
| Input | $2 | $4 |
| Output | 10 dollars | 20 dollars |
| Cache write operations | $2.50 | 5 dollars |
| Cache read operations | $0.20 | $0.20 |
The savings come from reduced token consumption. Sonnet 5.5 requires significantly fewer tokens for the same amount of work. In Anthropic's tests, this resulted in a task costing up to 30 percent less than with its predecessor. At low to medium workloads, the model outperformed Sonnet 5's best results in several tests, at about one-tenth the cost per task.
New protective mechanisms
Because its cybersecurity capabilities are roughly on par with Opus 5, according to Anthropic, Sonnet 5.5 is the first Sonnet model to launch with protection mechanisms like those found in the most powerful models. Routine tasks such as finding and fixing bugs in your own code remain possible. More risky security-related requests are noticeably handled by Sonnet 5.
Also new is protection against so-called distillation attacks. In these attacks, perpetrators attempt to exploit thousands of fake accounts to siphon off the capabilities of a model on a massive scale. The protection mechanisms in the biology domain remain unchanged compared to Sonnet 5.
When the middle model is sufficient
With Sonnet 5.5, the threshold at which Opus becomes necessary shifts significantly upwards. In professional and computer tests, both models perform almost identically, and Sonnet even has the edge when programming in the terminal – at half the price per token. We expect many developers to shift routine tasks to Sonnet and reserve Opus for planning and more complex decisions.
If you use the Claude app on iPhone, iPad, or Mac, the model runs there by default with medium effort. Even at this level, it surpasses the best result of Sonnet 5 in the terminal test. The same default setting applies in Claude Code, which can now test iOS apps directly in the simulator.
Will Sonnet be sufficient for your everyday use, or will Opus remain your standard model? Let us know in the comments which model you use for what.
