Claude Sonnet 5's API price rises 50% on September 1. That date is only one axis of the migration budget: Anthropic says its updated tokenizer can turn the same input into roughly 1.0 to 1.35 times as many tokens.

Those effects should not be added into an alarming “85% increase.” The rate change multiplies the bill; tokenizer sensitivity changes the quantity of input billed. Keeping them separate produces a useful range.

Four totals, not one

Anthropic's Sonnet 5 announcement lists introductory pricing through August 31 at $2 per million input tokens and $10 per million output tokens. Standard pricing is $3 and $15.

Take a workload previously counted as ten million input tokens and two million output tokens. Holding output fixed to isolate the documented input effect gives this range:

Period Input multiplier Input cost Output cost Total
Through Aug 31 1.00x $20.00 $20.00 $40.00
Through Aug 31 1.35x $27.00 $20.00 $47.00
From Sep 1 1.00x $30.00 $30.00 $60.00
From Sep 1 1.35x $40.50 $30.00 $70.50
A four-cell cost matrix showing the introductory and standard Sonnet 5 rates across the documented tokenizer range

The tokenizer footnote changes the migration test

Anthropic says the same input content may map to roughly 1.0–1.35 times as many tokens, depending on content. That range is too wide to hide inside a single nominal estimate. At the upper end, the example's input portion is $27 during the introductory period instead of $20, then $40.50 instead of $30.

The September increase remains exactly 50% for the same tokenized workload because both input and output rates rise by half. Token inflation is not another 35 percentage points on the whole bill; it is a sensitivity on input quantity.

A calendar and tokenizer gauge showing the two independent variables in a Sonnet 5 budget

Budget a range, then measure your content

This article does not test Sonnet 5 quality, latency, or tokenization. I did not buy access. The 1.0–1.35 range is Anthropic's documented estimate, and the totals are arithmetic built from it.

Once access is available, the useful migration test is straightforward: tokenize a representative sample of the actual prompts, retrieved context, tool results, and documents. Do not use a hand-picked short prompt if production inputs are code or long files. Then apply the rate in effect on the date the workload will run.

The announcement supplies an attractive single price. A defensible budget needs a date, an input mix, and a range.