16:40Machine-translated from Russian by AIRead in Azerbaijani

Anthropic Unveils Claude Sonnet 5.5 With Optimized Token Usage

Anthropic has unveiled the new Claude Sonnet 5.5 AI model, the second in its current lineup of version 5.5 neural networks. The new model is positioned as a balanced midrange solution combining high speed and cost efficiency for everyday tasks, programming, document analysis, and presentation generation.

Anthropic Claude Sonnet 5.5

The main change compared with Claude Sonnet 5 is a substantial increase in performance while reducing the amount of computation needed to complete typical tasks. By calling external tools more efficiently, the neural network takes fewer steps to process complex repositories, allowing it to compete successfully with more expensive AI systems. According to the developers, Sonnet 5.5 generates responses more than 30% faster than the previous version and costs up to 30% less in most scenarios, using significantly fewer tokens to do the same work.

In the benchmarks presented, Sonnet 5.5 handles most tasks significantly better than its predecessor. In a test of programming skills using the specialized Terminal-Bench 4.0 benchmark, the model scored 70.6%, demonstrating a huge lead over Sonnet 5, which scored only 10.3%. In the CursorBench 4.0 development environment, the new model achieved 55.5%, compared with 34.1% for Sonnet 5, coming close to the flagship Opus 5.5 at 57.8%. A similar pattern can be seen in comprehensive evaluations: in GDPval-AA, which assesses the completion of multidisciplinary professional tasks, the algorithm scored 1,844 points, virtually matching the flagship model’s 1,846 points. In personal computer control automation scenarios on the OSWorld 2.1 platform, it scored 80.1%, just 1.7% below the top model.

The company notes an overall reduction in real-world token usage, allowing complex work to be completed at a lower financial cost. At the same time, official pricing for access to the Claude Sonnet 5.5 API remains unchanged at $2 per million input tokens and $10 per million output tokens. The model has already been deployed and is available to users through the official website and Claude apps, as well as on the leading Amazon Web Services, Google Cloud, and Microsoft Azure cloud platforms.