Anthropic introduced the Claude Sonnet 5.5 model with improved performance

On September 28, Anthropic announced the Claude Sonnet 5.5 model, which operates more than 30% faster and costs 30% less per task compared to its predecessor, the Sonnet 5 model. The new model maintains prices at $2 per million input tokens and $10 per million output tokens. The Anthropic team reported that Sonnet 5.5 scored 70.6% on the Terminal-Bench 4.0 test, whereas Sonnet 5 scored only 10.3%, and Opus 5.5 – 66.4%.

Sonnet 5.5 became the second model in the Claude 5.5 family. On September 22, Anthropic also launched Opus 5.5, on the same day OpenAI released the GPT-6 Sol and Luna models. Anthropic noted that Sonnet 5.5 does not expand its capabilities, and its alignment review focuses on risks such as misleading users and aiding in high-stakes abuses. The model also includes cyber protection and anti-distillation, which prevents attempts to extract its logic for training competitive systems. The release of Claude Haiku 5.5 is expected in the coming weeks.

Nevertheless, the independent analytical company Artificial Analysis, after conducting testing, came to a different conclusion regarding the cost. According to their data, Sonnet 5.5 used significantly more tokens to achieve results. The analysis showed that under maximum load, Sonnet 5.5 used about 193,000 output tokens per task in the Intelligence index, which is 60% higher than Opus 5.5 and approximately 7 times more than GPT-6 Astra. However, clients with early access reported an opposite trend compared to Sonnet 5. For example, Balyasny Asset Management indicated that on their financial tasks, Sonnet 5.5 used about 121,000 tokens per response, while Sonnet 5 – 497,000. Artificial Analysis conducted testing of a preliminary version of the model with an error in structured output, which was corrected by Anthropic. The company plans to conduct relevant assessments again in the near future.