Anthropic cancels the Claude Sonnet 5 price rise it announced in June — $2/$10 pricing is now permanent
Anthropic confirmed on 11 August 2026 that it is scrapping the planned 1 September price increase for Claude Sonnet 5, locking introductory pricing of $2/$10 per million input/output tokens in permanently instead of reverting to the previously announced $3/$15.
25 August 2026
When Claude Sonnet 5 launched at the end of June, Anthropic set introductory pricing of $2 per million input tokens and $10 per million output tokens, due to expire on 31 August and revert to standard pricing of $3/$15 from 1 September — a flat 50% increase. On 11 August, Anthropic reversed that decision: the $2/$10 rate is now permanent, with no expiry date attached.
It’s a genuinely unusual move. Model pricing in 2026 has mostly gone one direction — up, as labs try to recoup the compute cost of frontier training runs — or down only when a competitor forces the issue, as OpenAI and Google have done repeatedly this year in a running price war. Anthropic cancelling its own planned increase, rather than being pushed into a cut, reads as a signal that Sonnet 5 adoption and retention matter more to Anthropic right now than near-term margin on the token bill.
Why the sticker price was never the whole story
Sonnet 5 shipped with a new tokenizer that generates meaningfully more tokens for the same input than its predecessor did — for some workloads, around 30% more. That inflation happens regardless of the headline rate. So while cancelling the 1 September increase removes the sharpest edge of the looming cost jump, teams running significant Claude Code or API volume should still expect their effective spend to have risen since June, just by less than the worst-case scenario would have produced. The rate is stable now; the token-count assumption underneath it still deserves a periodic check.
So what
For anyone budgeting an AI-assisted build, this is good news with a caveat attached: Anthropic has removed the single biggest planned cost variable for the rest of the year, which makes multi-month project budgets easier to hold to. It’s not a reason to stop tracking actual token consumption against estimates — the tokenizer change means the honest comparison is real usage data, not the published rate card. We build model and token economics into project scoping as standard, rather than treating AI usage costs as a fixed line item carried over from the kickoff estimate. See how that fits into AI-assisted development, or get in touch to talk through a build.