Anthropic introduced Claude Sonnet 5.5. It is more efficient, faster, and often matches Opus 5.5.

Less than a week after the release of Opus 5.5, Anthropic is complementing the new series with its cheaper and faster sibling. Claude Sonnet 5.5 is aimed at well-defined daily tasks, such as fixing bugs in code or creating documents, presentations, and spreadsheets. If you use Claude via API, you will pay the same for tokens as with Sonnet 5, but the new model will consume fewer of them for the same task.

Often matches Opus 5.5 in tests

Anthropic states in its announcement that the model has improved in all areas compared to its predecessor, dramatically in some. The most significant leap is reported from the Terminal-Bench 4.0 test, which examines complex command-line tasks. According to the company’s table, Sonnet 5.5 achieved 70.6% in it, while Sonnet 5 only reached 10.3%. It even slightly outperforms Opus 5.5 with 66.4% here.

In the GDPval-AA test, where models perform real-world tasks from 44 professions, it lags behind Opus 5.5 by just two points (1,844 vs. 1,846) and surpasses Sonnet 5 by approximately 400. OpenAI’s competing GPT-6 Sol has 1,487 points in the table. The new model also remains close behind Opus 5.5 in computer control and graph reading. Furthermore, it is the first Sonnet to complete Pokémon Red based solely on screenshots. Despite all this, the company admits that for complex and open-ended tasks requiring long-term judgment, Opus 5.5 is still distinctly stronger, which is logical given the long-term definition of the Claude model series.

According to the company, the improvement is most noticeable in programming. In the FrontierCode test, Sonnet 5.5, with the same settings, has a score 10 points higher than Sonnet 5 at roughly one-fifteenth of the cost. Testers liked how quickly it familiarizes itself with foreign code and that it more often combines tool calls, allowing it to complete tasks in fewer steps. Epic Games COO Daniel Vogel states that the model handled tens of thousands of lines of game architecture code and multi-hour tasks.

The company also highlights its flair for design. The new Sonnet is said to provide better refinement to application aesthetics and create a presentation from a template that requires only minor adjustments. In an internal test, it was given the quarterly results of a publicly traded company along with transcripts of investor calls and was tasked with preparing a ten-slide presentation based on a template. Two experts deemed its first version ready for submission.

Sonnet 5.5 is much more efficient at the same price

Developers will pay approximately 43 CZK (2 dollars) per million input tokens and 214 CZK (10 dollars) per million output tokens, which is half the cost of Opus 5.5. Reading from cache costs approximately 4 CZK per million tokens. The savings, therefore, do not come from a lower price list but from efficiency: according to Anthropic, the model requires fewer tokens for the same work, and thus the task will cost up to 30% less than with Sonnet 5. Slack, for example, measured approximately 14% fewer output tokens for its Slackbot.

Furthermore, it generates text over 30% faster, making it the fastest Sonnet in history. You can balance speed and cost by choosing the effort level. In Claude applications and Claude Code, the default is medium effort, while on the developer platform, it’s high. With the medium effort level, Sonnet 5.5 significantly surpassed Sonnet 5’s best result in Terminal-Bench, according to the company, at less than a tenth of the cost.

The model is available now on all platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure clouds, even in data-retention-free mode. The 5.5 series will be concluded in the coming weeks by Haiku 5.5, designed for low-cost, high-volume deployment.

New Sonnet receives stronger security safeguards for the first time

According to Anthropic, its cybersecurity capabilities are on par with Opus 5, making it the first Sonnet to receive similar safeguards as the company’s most powerful models. Searching for and fixing bugs in its own code will continue, but riskier security tasks will visibly shift to the older Sonnet 5. Protection in the field of biology remains the same as its predecessor, but the company notes it may inadvertently capture some queries from microbiology and virology.

The new Sonnet also defends against so-called distillation, where attackers use thousands of fake accounts to extract the model’s capabilities for their own AI without security restrictions. Therefore, it is the first Sonnet to receive classifiers that prevent the exploitation of its reasoning. The model’s reasoning is now tied to the account that created it, which you will notice if you transfer conversations between accounts, for example, when switching accounts in the middle of working in Claude Code. In an automated behavior audit with approximately 1,850 scenarios, it performs as well as or better than Sonnet 5, according to the company, and among all its models, it least tests the boundaries of the environment in which it runs.

Is Sonnet 5.5 sufficient for your everyday work, or would you gladly pay extra for Opus?

Source: Anthropic