Claude Sonnet 5 ships with 1M token context and new tokenizer that increases costs 30-40% for English text

Anthropic released Claude Sonnet 5 today with a 1 million token context window and 128,000 maximum output tokens. The model claims performance close to Opus 4.8 at lower nominal prices, but a tokenizer change substantially increases effective costs.

Key specifications

• Context window: 1 million tokens
• Maximum output: 128,000 tokens
• Pricing: $3 per million input tokens, $15 per million output tokens (introductory discount to $2/$10 through August 31)
• Same pricing as: Claude Sonnet 4.6
• Adaptive thinking: Enabled by default unless explicitly disabled via "thinking": {type: "disabled"}

Breaking changes

The model removes support for traditional sampling parameters: temperature, top_p, and top_k are no longer available in the API.

More significantly, Sonnet 5 introduces a new tokenizer that generates approximately 30% more tokens than Sonnet 4.6 for identical input text. Testing on the Universal Declaration of Human Rights showed:

• English text: 1.42x more tokens (42% increase)
• Spanish text: 1.33x more tokens (33% increase)
• Python code: 1.27-1.28x more tokens (27-28% increase)
• Simplified Mandarin Chinese: 1.01x more tokens (effectively unchanged)

For a 4,279-line Python file, Sonnet 4.6 generated 44,014 tokens while Sonnet 5 generated 56,113 tokens—a 27% increase that translates directly to higher costs despite identical per-token pricing.

Safety and capabilities

According to Anthropic's system card, "Sonnet 5 is significantly less capable at cyber tasks than Mythos 5: its safeguards are thus similar to those we apply to Opus 4.7 and Opus 4.8 (models that are more capable than Sonnet 5 but much less capable than Mythos 5)." This reduced capability in cybersecurity tasks enabled the model's release without U.S. government regulatory blocks.

What this means

While Anthropic maintains the same nominal pricing as Sonnet 4.6, the new tokenizer makes Sonnet 5 effectively 27-42% more expensive for most use cases involving English text and code. Only users working primarily with Chinese text will see pricing parity. The removal of sampling parameters may require developers to adjust their existing implementations. The 1 million token context window represents a substantial increase from previous models, though developers should account for the higher tokenization rate when planning context budgets.