Your prompt did not change and costs 30% more

On 16 April 2026 the text of your prompt stayed the same and its bill did not. What changed was the unit it is measured in: the Claude Opus 4.7 tokenizer turns the same text into roughly 30% more tokens.

What has shipped

claude-opus-4-7, on 16 April 2026, at the same price as its predecessor: $5 per million input tokens and $25 per million output. Three changes, and the one that was announced least is the one you feel most.

The tokenizer. From Opus 4.7 onward, the same text produces roughly 30% more tokens. The price per token did not move; the number of tokens did. A 3,000-token prompt becomes about 3,900 without anyone editing a line.

The xhigh level, which slots between high and max and is aimed at long agentic and coding runs: thirty minutes or more, with million-token budgets.

High-resolution vision: the long edge of an image goes from 1,568 to 2,576 pixels and can cost up to about three times the image tokens.

And it did not stay on that model. Sonnet 5, on 30 June 2026, carries the same tokenizer and the same 30%.

What changes for reasoning

Nothing breaks here, and that is the problem. The prompt works, the tests pass, the answers are the same. What changes lives in two places nobody reads in a git diff: the cost line and the point where the prompt stopped fitting.

The trap is that the failure arrives late. A prompt that went over the context window used to return an error; one that is now 30% over gets truncated, or loses its headroom for the reply. Neither carries the tokenizer's name in the message.

The other two changes push in the same direction. xhigh is meant for thirty-minute runs with million-token budgets, which are exactly the runs that notice the 30% most. Switching it on by default adds depth on top of a cost that already rose, and high-resolution images add up to three times their tokens on top of the same 30%.

What changes for writing prompts

Three numbers and a date. You measure the fixed prefix —system block and tools—, a typical turn and a typical output, with the provider's token-counting API rather than a rule of thumb. You compare against what you wrote down before 16 April and divide: that is your prompt's real multiplier, and it is almost never exactly 1.30.

With the number in front of you, three things get re-checked.

Context headroom. What fits has not changed; the unit has. A prompt that left 15,000 tokens free now leaves less, and the room for the reply narrows without anyone touching it.

Cache thresholds. Cache is priced per token, so the same prefix costs roughly 30% more to write and to read. And any minimum expressed in tokens is measured in a unit that just changed: the minimum cacheable prompt drops to 1,024 tokens with Opus 4.8 on 28 May 2026, and a prefix that sat near the threshold may have changed sides without anyone noticing.

Price per call. The price per million did not move. The call did: every token of that text costs what it cost before, and there are more of them.

cost.md, next to the prompt, measured with the Opus 4.7 tokenizer:

fixed prefix (system + tools)   [tokens]
typical turn                    [tokens]
typical output                  [tokens]
cost = (input × 5 + output × 25) / 1,000,000
measured on [date] · re-check when the model or the tokenizer changes

What to stop doing

• Shortening the prompt blind to square the bill. The 30% is not in the prose, it is in the tokenizer. Cutting removes instructions and leaves the multiplier.
• Counting tokens with a rule of thumb. If your code estimates with characters per token, that rule was written with the old tokenizer.
• Trusting a budget measured before 16 April. That goes for context, for cache and for price.
• Switching to xhigh by default. It is the level for runs of thirty minutes or more; on a short call you pay for depth you do not use.
• Turning high resolution on across the whole flow. High-resolution images are up to about three times the image tokens, on top of the 30% you already carry.
• Re-measuring once and calling it done. The same tokenizer reached Sonnet 5 on 30 June; the number is per model, not per vendor.

Sources

• Release notes
• Claude Opus 4.7
• Effort