Anthropic's new flagship tier model, Claude Sonnet 5, keeps the same list price as its predecessor, but a redesigned tokenizer inflates token counts by 28–42% depending on language. Existing integrations also lose support for three sampling parameters.
Anthropic released Claude Sonnet 5 on June 30, 2026, billing it as a mid-tier model that closes most of the gap with its flagship Opus 4.8 while keeping the same list price as the outgoing Sonnet 4.6. That second clause is true. It is also misleading in dollar terms, because the new model ships with a redesigned tokenizer that inflates how many tokens each prompt and completion actually consumes. For a team currently running on Sonnet 4.6, "same price" translates to a meaningful increase in the line item at the end of the month.
Independent AI reviewer Simon Willison benchmarked the new tokenizer against Sonnet 4.6 across four languages and found a consistent step-up in token count per unit of text: English prompts run roughly 1.42× longer, Spanish about 1.33×, Python source code about 1.28×, and Simplified Mandarin close to flat at 1.0×. (Simon Willison — What's new in Claude Sonnet 5) Anthropic's own announcement confirms the tokenizer change as the underlying mechanism without publishing the per-language ratios. (Anthropic — Introducing Claude Sonnet 5)
The math carries over directly to billing. Sonnet 5 lists at $3 per million input tokens and $15 per million output tokens, with an introductory discount of $2 and $10 running through August 31, 2026. (Anthropic — What's new in Claude Sonnet 5 (developer docs)) At those rates, an English workload that produced one million input tokens on Sonnet 4.6 produces roughly 1.42 million on Sonnet 5, and the bill moves with it. The headline reads "no price change"; the invoice tells a different story.
Beyond the tokenizer shift, the developer-facing delta list is unusually long for what Anthropic is positioning as a routine model bump. Sampling parameters temperature, top_p, and top_k are no longer supported; code paths that set those values will need to drop them. Context window expands to 1,000,000 tokens with a maximum output of 128,000 tokens. Adaptive thinking is on by default and must be explicitly disabled with thinking: {type: "disabled"} if a workload requires the old behavior. (Anthropic — What's new in Claude Sonnet 5 (developer docs))
Anthropic frames Sonnet 5's capability level as "close to Opus 4.8 at lower prices," with the same tool and platform feature set as Sonnet 4.6 underneath. (Anthropic — Introducing Claude Sonnet 5) That positioning matters for any team deciding whether to stay on the mid tier or step up to Opus: Sonnet 5 narrows the gap rather than widening it.
There is also a regulatory dimension, which Anthropic surfaces in the system card and Simon Willison pulls forward: Sonnet 5 is "significantly less capable at cyber tasks" than Mythos 5, an apparently gated reference model that has not been released. Safeguards applied to Sonnet 5 match those used for Opus 4.7 and Opus 4.8, and Anthropic concludes the model was releasable without US government intervention. (Simon Willison — What's new in Claude Sonnet 5) That framing is Anthropic's own, not third-party validated; treat the cyber-capability claim as company self-reporting rather than a measured benchmark.
What this means for a team currently running Sonnet 4.6 is concrete. Recalc the cost line assuming 30–40% more input tokens for English-language workloads, update prompt code that sets temperature, top_p, or top_k, and decide whether adaptive thinking on by default fits the use case before flipping the model in production. The introductory price window closes August 31, 2026, so any meaningful migration test should be done before that date rather than after.