Claude Sonnet 5's introductory API pricing -- $2 per million input tokens and $10 per million output tokens -- became Anthropic's standard, permanent rate for the model on August 10, according to the company's own developer changelog. The single-line entry canceled a 50% increase to $3 and $15 per million tokens that had been on the calendar since the model's June 30 launch and was scheduled to take effect September 1 -- one day after Sonnet 5's original introductory window was set to close.
Sonnet 5 API pricing, planned vs. actual
- Input tokens
- Output tokens
- Status
The increase had been Anthropic's stated plan for weeks -- a pricing page that, as recently as August 1, still showed the $3/$15 rate taking effect September 1, before the August 10 changelog entry reversed it. Anthropic has not published a separate blog post or statement explaining the change; the changelog entry is the only public record of it.
One cost driver the reversal doesn't touch: Sonnet 5 launched with a new tokenizer that Anthropic's own release notes say produces "approximately 30% more tokens" than the prior tokenizer for the same text, with the exact increase depending on content and workload shape. That change was priced in from day one and has nothing to do with the per-token rate -- it means the same prompt was already billing as more tokens on Sonnet 5 than it would have on Sonnet 4.6, independent of whatever Anthropic charges per token.
The model itself hasn't changed alongside the pricing news. Sonnet 5 launched June 30 with a 1-million-token context window and a 128,000-token maximum output, the same platform features as Sonnet 4.6 with adaptive thinking on by default -- specs that were already priced into the original $2/$10 introductory rate and remain priced into it now that the rate is permanent rather than temporary.
- Claude Sonnet 5's $2/$10 per-million-token launch pricing became Anthropic's permanent rate on August 10.
- The change cancels a scheduled 50% price increase to $3/$15 that was due September 1.
- The increase had been on the calendar, unchanged, since the model's June 30 launch.
- Anthropic has not published a separate explanation beyond the one-line changelog entry.
- Caveat: a separate tokenizer change already bills roughly 20-35% more tokens per request, untouched by this reversal.