Anthropic shipped Claude Sonnet 5 on June 30, calling it the most agentic Sonnet it has built, and priced it to move: an introductory $2 per million input tokens and $10 per million output, holding through August 31 before it steps up to $3 and $15. Anthropic says it runs close to Opus 4.8 at roughly a third of the cost, and the framing everywhere was the obvious one. The cost of running agents just dropped, and a price war is starting. Then you read the pricing note, and the discount gets more complicated.
I've carried a bag for usage-priced software, and the pattern is always the same. The number that closes the deal is the per-unit price on the slide. The number that decides whether the customer renews is what they actually spent to get a job done. Those two drift apart the moment the unit changes definition, and that is exactly what happened here.
The sticker fell. The unit didn't hold still.
Sonnet 5 ships with a new tokenizer, the component that chops your text into the tokens you're billed on. Anthropic says the same input can map to roughly 1.0 to 1.35 times as many tokens as before, and it was candid about the consequence: it set the introductory price so the transition would be, in its words, roughly cost-neutral. Look at the mechanics. A headline of $2 against the old $3 reads as a third off. If the identical task now generates a quarter to a third more tokens, part of that cut is arithmetic rather than savings. For a light workload you may still come out ahead. For a heavy agentic loop that reads long documents and reasons in many steps, the exact use Sonnet 5 is sold for, the token inflation lands hardest, right where the model is meant to shine.
Why this bites during a price war
It matters more because everyone is now shopping agents on price. When vendors compete on dollars per million tokens, the comparison only holds if a token means the same thing across models, and it doesn't. Each lab tokenizes differently, and now Anthropic has changed its own mid-stream. A buyer comparing Sonnet 5's $2 to another model's $2.50 on the spec sheet can pick the option that costs more per finished task and never see it, because the cheaper-looking model quietly needs more tokens to do the same work. The pricing page stopped being a fair scoreboard.
If you sell or buy AI on tokens
For anyone reselling agentic workflows, this is a margin question you should answer before your customer does. If you quoted a client based on old token math, run the new tokenizer against your real traffic before you celebrate the cut, because your cost of goods may not have moved the way the announcement implies. For anyone buying, add one line to your evaluation: cost per completed task on your own workload, measured through each model end to end, not list price per token. Take a representative job, the document your team actually processes, the agent loop you actually run, and read the bill at the bottom. That is the only number that survives a tokenizer change.
None of this makes Sonnet 5 a bad deal. For plenty of workloads it's a real cut, and cheaper capable agents are good for everyone building on them. The point is narrower, and it's about how you count. Model vendors will compete loudly on per-token price all year, and the unit under that price is theirs to redefine. Keep your own meter running on cost per outcome, or you'll re-quote your customers off a discount your invoice never delivered.