- Published on
What's new in Claude Sonnet 5
- Authors
- Name
- DeepHorus
Anthropic released Claude Sonnet 5 on June 30, 2026, positioning it as a model that offers performance close to the more expensive Opus 4.8 at significantly lower prices. The announcement, detailed in the developer documentation, highlights key changes in API parameters, tokenization, and safety measures that differentiate it from its predecessor, Sonnet 4.6, and from higher-tier models like Mythos 5.
Unlike previous Sonnet models, Sonnet 5 no longer supports sampling parameters such as temperature, top_p, and top_k, simplifying the API interface. It retains a 1 million token context window and extends maximum output tokens to 128,000, matching the capabilities of Opus-level models. Additionally, adaptive thinking is enabled by default, though developers can disable it with a specific parameter.
The pricing structure remains unchanged from Sonnet 4.6 at $3 per million input tokens and $15 per million output tokens, with an introductory discount to $2 and $10 respectively until August 31, 2026. However, a new tokenizer increases the effective cost by approximately 30% for English text, as the same input now generates about 30% more tokens. Testing with Claude Token Counter revealed a 1.4x increase for English, 1.33x for Spanish, 1.28x for Python code, while Simplified Mandarin saw no significant change.
Anthropic addressed regulatory concerns by ensuring Sonnet 5 is significantly less capable at cyber tasks than Mythos 5, allowing for similar safeguards as applied to Opus 4.7 and 4.8. This decision likely helped the company avoid restrictions from the U.S. government, which has scrutinized advanced AI models for potential dual-use risks.
The introduction of a new tokenizer has drawn mixed reactions from developers, as the 30% effective price hike offsets the lower base pricing. While Sonnet 5 offers improved performance and capabilities, the increased cost may impact adoption among cost-sensitive users. Some developers have noted that the model's ability to handle longer outputs and larger contexts could justify the expense.
Looking ahead, Anthropic plans to maintain competitive pricing while continuing to refine its tokenizer efficiency. The company remains focused on balancing capability and safety, as evidenced by the deliberate limitations placed on Sonnet 5's cyber performance. As the AI landscape evolves, Sonnet 5 sets a new benchmark for mid-tier models, but its long-term success will depend on developer acceptance of the revised pricing model.
