Claude Sonnet 5 Emits 30% More Tokens Than Sonnet 4.6

Anthropic shipped Claude Sonnet 5 on June 30, calling it a drop-in replacement for Sonnet 4.6, per Simon Willison's read of the platform docs.

The new tokenizer emits roughly 30% more tokens for the same input text. Token budgets, context limits, and cost projections built against Sonnet 4.6 are already off by nearly a third.

Manual extended thinking now returns HTTP 400. So does setting temperature, top_p, or top_k to anything but the default, since adaptive thinking runs by default now. These calls don't degrade quietly, they fail outright.

Sonnet 5 is the first model at this tier with real-time cybersecurity safeguards. A refusal now comes back as HTTP 200 with a stop_reason of refusal. Error handling that only checks status codes will read that refusal as a normal completion and pass along an empty response.

Yes, the pricing is good: $2/$10 intro rates through August 31, a 1M-token context window, and 128k max output. There's no Priority Tier, which matters if you were counting on it.

None of that makes this a drop-in. Recount your token budgets and add a check for a refusal stop_reason before you flip the switch.