The launch price for Claude Sonnet 5 ($2 input, $10 output per million tokens) ends on 31 August 2026. On 1 September, the standard price takes over: $3 and $15. That's a net increase of 50% on both lines. If you've built a budget on current prices, it's time to redo the maths.
There's a category of unpleasant surprise that's particularly annoying: the one you could have seen coming. The end of Sonnet 5's promotional price falls into that category, and it hides a second effect that most announcements gloss over. Here's the full calculation.
What changes on 1 September
Anthropic launched Sonnet 5 at the end of June with an introductory price, as we noted in our article on its launch. The mechanics are simple.
| Per million tokens | Until 31 August | From 1 September |
|---|---|---|
| Input | $2 | $3 (+50%) |
| Output | $10 | $15 (+50%) |
| Cache read | $0.20 | $0.30 |
No automatic extension is planned. It's a firm deadline.
The trap: the price doesn't change, the bill does
Here's the element many will discover in September. At $3 and $15, Sonnet 5 lands exactly on the price of its predecessor, Sonnet 4.6. On paper, the price list is therefore identical. A finance manager comparing the two lines will see no anomaly.
Except Sonnet 5 ships a new tokenizer. And a different tokenizer splits the same text into a different number of tokens. According to measurements reported by several analyses, Sonnet 5 would count around 30% more tokens for identical text, with gaps of up to 27% on Python code and 42% on English prose.
If you've followed our article on the cost of French, the mechanism will be familiar: the tokenizer is the tool that splits text into billable units, and finer splitting means more tokens for the same content. The unit price is identical, the billed quantity goes up.
Add the two effects together and the result becomes clear: from 1 September, an unchanged workload could cost around 35% more than the same task on Sonnet 4.6, even though the price list shows exactly the same figures. The 50% price increase is visible. The token inflation isn't. It's the combination of the two that catches you out.
The third factor: effort level
One last element deserves your attention. Sonnet 5 offers adjustable effort levels, and adaptive thinking is on by default. As we showed with Kimi K3 and its reasoning tokens, a model that thinks at length consumes a huge amount.
Two situations make Sonnet 5 lose its pricing advantage. The first: pushing it to maximum effort to match Opus 4.8 on hard tasks, where the volume of reasoning tokens can exceed Opus's cost at comparable accuracy. The second: sending a huge context on every call when a fraction would suffice, since the tokenizer then multiplies those input tokens.
The remedy is explicit in both cases: rather than pushing Sonnet to its limits, escalate to Opus 4.8 for genuinely hard tasks, and trim your context ruthlessly.
What you can do before 31 August
Measure now, not in September. Run a representative sample of your real traffic and record your actual token consumption per task. That's the only figure that matters, and you'll only get it by measuring on your own setup.
Budget at September prices. Model your costs at $3 and $15, with your measured token multiplier. Building a back-to-school budget on the promotional price guarantees an unpleasant surprise.
Set up caching now. If a large part of your context is stable from call to call, like a long system prompt or a codebase, caching can sharply reduce the cost of repeated input tokens. Well built, this mechanism can offset a good chunk of the September increase.
Bring forward what can be brought forward. Batch processing, evaluation campaigns, history re-runs: anything that isn't urgent but will need doing anyway is better executed before 31 August. It's a limited-time discount, so you might as well use it.
What to remember
Sonnet 5 remains an excellent deal for agentic work, with performance close to Opus 4.8 for a fraction of the price. The problem isn't the model, it's the shape of the curve: migrating in July lowers the bill, nobody worries, then on 1 September the price jumps 50% on exactly the same traffic, with tokens already inflated by the new tokenizer.
The lesson goes beyond this specific case. A listed price tells you nothing about your real cost until you've measured how many tokens your work actually consumes. The teams that will avoid the September surprise are those doing that calculation in August, not those discovering their bill in October.