Source: Simon Willison — 2026-06-30
Summary
In his technical breakdown of Anthropic's Claude Sonnet 5 launch, Simon Willison digs past the official announcement into the developer documentation and finds a consequential detail: a new tokenizer makes English-language input roughly 1.4x more expensive in token count, even though per-token pricing is unchanged. He notes the effect is smaller for other languages, meaning the real-world cost impact varies significantly by workload language mix.
Key Takeaways
- The tokenizer change means effective cost-per-request can rise ~1.4x for English input despite no headline price change — a detail easy to miss without reading the docs closely.
- Non-English languages see a smaller effect, which shifts the relative cost economics of multilingual agentic workloads.
- A reminder that "unchanged pricing" announcements can hide real cost deltas buried in tokenizer or context-handling changes.
- Useful for anyone budgeting or forecasting spend on agentic workloads built on the new model.