Claude Sonnet 5's price jumps 50% on September 1 — the tokenizer makes it worse

Anthropic's introductory pricing for Claude Sonnet 5 ends August 31, 2026. Starting September 1, the standard rate of $3 per million input tokens and $15 per million output tokens replaces the launch rate of $2/$10 — a 50% jump on paper, according to Anthropic's own [pricing documentation](https://platform.claude.com/docs/en/about-claude/models/whats-new-sonnet-5).

That's the headline. It is not the number that matters most.

The part that doesn't show up in the price sheet

Claude Sonnet 5 shipped with a new tokenizer. Anthropic says the same input text now produces roughly 30% more tokens than it did on Claude Sonnet 4.6 — the exact figure depends on content, but 30% is the number in their own migration guide. Anthropic set the introductory $2/$10 rate specifically to make the Sonnet 4.6 → Sonnet 5 switch close to cost-neutral during the promotional window. That offsetting math disappears on September 1.

Stack the two changes and you're not looking at a 50% bill increase. You're looking at roughly 50% higher per-token pricing applied to a token count that's already ~30% higher for identical text. Run a workflow that cost $100/month on Sonnet 4.6, and if you migrated during the intro window without re-measuring, September's bill for the same workload is not $150. It compounds toward something closer to $190–200, depending on how much of your spend is input versus output tokens.

There's a second multiplier most teams haven't accounted for: adaptive thinking is on by default on Sonnet 5. Requests that ran with no thinking at all on Sonnet 4.6 now run with thinking turned on unless you explicitly disable it. Thinking tokens count against `max_tokens` and against your bill. If your automation calls Claude without an explicit `thinking: {type: "disabled"}` flag, you inherited this cost silently the moment you upgraded the model ID.

Why this matters if you're running Claude in production

Any business running agents, RAG pipelines, support bots, or scheduled automation on Claude has three things happening at once, not one:

• **Per-token price**: up 50% after August 31.
• **Tokens per request**: up ~30% from the tokenizer change, same input text.
• **Tokens per request, again**: up further if adaptive thinking is silently running on calls that used to skip it.

None of these show up if you're just glancing at the per-million-token rate card. They show up in the invoice, three weeks after you've already shipped the migration and moved on to the next thing.

The honest caveat

This isn't a reason to panic or to switch providers reflexively. A few things cut the other way:

Sonnet 5 is a genuine capability upgrade over 4.6, particularly on coding and agentic tasks, at the same headline per-token price. If it solves a task in fewer retries, fewer follow-up turns, or a shorter agent loop, the cost-per-completed-task can still come out flat or lower even though cost-per-token is up. Token count and dollar cost are not the same metric, and only one of them is what you actually care about.

Enterprise volume pricing through AWS Bedrock, Google Vertex, or Microsoft Foundry may already differ from the list-price numbers above — if you're on a negotiated rate, recheck that contract rather than assuming the public rate card applies. And if your workloads are light — a few thousand calls a month, not millions — the absolute dollar swing may be small enough not to be worth engineering time to chase.

What doesn't cut the other way: assuming "same price, better model" from the launch announcement means your bill won't move. It will. The per-token rate card is not the same thing as the per-task cost, and the gap between the two just got wider.

What to do before August 31

Don't wait for the September invoice to find out what changed. Two things, both doable in an afternoon:

1. **Recount your actual prompts** against Claude Sonnet 5 using Anthropic's token counting endpoint, not against your old Sonnet 4.6 numbers. Do this for your highest-volume automation first — that's where the compounding hits hardest. 2. **Check every Claude call in your codebase for an explicit thinking setting.** If none is set, you're running adaptive thinking by default on every request. Decide deliberately whether that's worth the tokens for each workflow, rather than inheriting it by omission.

If you don't have visibility into what your current Claude spend actually is by workflow — not by month, by workflow — that's the more urgent gap. Pricing changes are predictable and announced in advance. You can only get ahead of them if you already know what you're spending today.