Claude Sonnet 5.5 speeds up AI and cuts costs

Claude Sonnet 5.5 is now available as a faster, more efficient version of Anthropic's model. The company says it generates responses more than 30% faster and can cost up to 30% less per task than Sonnet 5, while keeping the same per-token price.

It is the second model in the Claude 5.5 family, after Opus 5.5. Its goal is not to solve the most open-ended and complex problems, but to handle everyday tasks better: coding, debugging, preparing documents, creating presentations, and working with spreadsheets.

More coding capability

The most visible improvement appears in coding. In Terminal-Bench 4.0, a test that measures whether a model can work as an agent inside a terminal environment, Sonnet 5.5 scored 70.6%, compared with 10.3% for Sonnet 5.

In FrontierCode, another evaluation of real-world coding tasks, it reached up to 52.1% at the highest effort level. Opus 5.5 remains ahead in this test, with 54.4%, but the gap has narrowed significantly.

In practice, this means Sonnet 5.5 can understand a codebase, locate a bug, and propose changes in fewer steps. According to Anthropic, early users also found that the model groups tool calls more effectively, reducing both time and usage.

It is not just about getting more answers right. If it needs fewer operations to complete a task, it can also be cheaper for projects that use AI intensively.

A model for everyday work

Sonnet 5.5 also improves at knowledge tasks and computer use. In GDPval-AA, an evaluation covering jobs from 44 professions and nine industries, it scored 1844 points, nearly matching Opus 5.5, which scored 1846. Sonnet 5 reached 1449.

Anthropic says the model performs especially well at:

• Creating documents, presentations, and spreadsheets with less follow-up editing.
• Following visual templates and improving interface design.
• Analyzing charts and images.
• Maintaining clearer work conversations.
• Completing long, multi-step tasks.

In an internal test, the company gave it financial materials and transcripts from an earnings call, along with a slide template. The model created a 10-slide presentation that two experts considered ready to send without changes.

There is also a notable detail: Anthropic says Sonnet 5.5 was the first Sonnet model able to beat Pokémon Red using only screenshots as guidance. This demonstrates visual understanding and planning, not a promise that the model will play like a person in any video game.

The price stays the same, but the real cost may fall

The official price per million tokens, the text units the model processes, remains:

• 2 dollars per million input tokens.
• 10 dollars per million output tokens.
• 0.20 dollars per million tokens read from cache.

The difference is that Sonnet 5.5 typically needs fewer tokens to do the same work. That is why Anthropic estimates that the cost per task can be up to 30% lower than Sonnet 5. The savings will depend on the task, the selected effort level, and how much text the model generates.

In Claude Code and Anthropic's applications, the default level is medium. On the developer platform, it is high. A lower level prioritizes speed and cost, while a higher one makes the model spend more time reasoning and reviewing its answer.

What this changes for you

If you use Claude to write, code, or prepare work materials, Sonnet 5.5 aims to become the balanced option for price, speed, and quality. It should respond faster and use fewer resources for routine tasks, while Opus 5.5 remains reserved for open-ended work that requires more judgment and sustained reasoning.

The model is now available on Anthropic's platform and on Amazon Web Services, Google Cloud, and Microsoft Azure. Developers can access it using the identifier claude-sonnet-5-5.

Anthropic is also adding new protections. Because its cybersecurity capabilities are approaching those of more advanced models, Sonnet 5.5 includes safeguards for high-risk requests. Normal software development should not be affected, but some cybersecurity tasks may be routed to Sonnet 5.

The company also says the model matches or improves on Sonnet 5 in most of its automated alignment audits, conducted across about 1,850 situations. Even so, it acknowledges that no evaluation detects every possible failure.

The important signal is not that Sonnet 5.5 replaces the most capable models, but that it brings some of their performance to a faster, cheaper model. If the results hold outside benchmarks, the competition will increasingly come down to a practical question: how much useful work can AI complete for every dollar and every second?