Keep model budgets honest: check price date and token multiplier

Expired prices are the quiet budget leak

Your model-cost spreadsheet still has a launch-week price row. When the project outlived the promotion, that row is now a date, not a rate.

Claude Sonnet 5 began on June 30, 2026 at $2 per million input tokens and $10 per million output tokens, with that rate labeled introductory through August 31, 2026.

The planned increase after the introductory period would have raised both rates by 50%, from $2/$10 to $3/$15 per million tokens.

On August 10, 2026, Anthropic changed the $2/$10 introductory rate into the standard rate and removed that planned September 1, 2026 move.

For a side project, the market story matters less than the spreadsheet error: a row that says introductory has an end date, and keeping it after that date builds the budget on a price that no longer exists. For future bills, replace the expired introductory row with the current standard rate. The leak is small enough to miss and large enough to distort a weekend project's total.

The rate is not the billable count

The price per million tokens can look stable, but the billable token count can change when the model changes. The billable count is what the invoice multiplies by the rate, not the raw character count or your draft token estimate.

Sonnet 5's newer tokenizer, which converts text into tokens, produces about 30% more tokens for the same text than the tokenizer used by Sonnet 4.6 and earlier models.

Anthropic characterizes that effect as a token-count multiplier between about 1.0x and 1.35x, with the exact value depending on content type.

The range matters because a raw token estimate can be wrong in both directions. A plain-text prompt may sit near the lower value, while a content-heavy prompt may sit closer to the higher value. Your budget needs a multiplier, not just a guess.

When comparing models, the listed price can be misleading. Sonnet 5's standard $2/$10 rate is lower than Sonnet 4.6's $3/$15 rate, so the newer model can cost less per million tokens before you count the tokenizer.

But the billable count may still rise. A cheaper rate and a larger token count can offset each other, so the budget check has a price part and a count part.

The budget check belongs before you ship

The check is small enough to run before a weekend deploy and small enough to keep in a spreadsheet note.

Each line is testable in a minute, and the order matters.

• Delete any price row with an expiry date.
• Multiply expected tokens by the model's tokenizer multiplier before applying the current standard rate.

The rate line keeps the price current. The count line keeps the billable number honest.

For a project that uses the Batch API, keep that rate separate. Sonnet 5's Batch API rate stays at $1 per million input tokens and $5 per million output tokens, a 50% discount from the standard $2/$10 rate.

A small spreadsheet is easier to trust

A side project spreadsheet does not need many columns. It needs a place for the rate, the expected input count, the expected output count, the multiplier, and the date you last checked the rate. Keep the date visible. A rate without a check date is just a number. If the date is older than your last deploy, treat the row as stale.

The price row should hold the standard rate, not the launch rate. A live promotion gets its end date next to the rate. An ended promotion gets replaced by the current standard rate.

The multiplier is where the model change appears. A content-heavy prompt may need a higher value than a plain-text prompt. When unsure, use the upper end of the stated range for that model.

That value feeds the billable count. The result is the number your budget should use. It is less pretty than a raw count, but it is closer to the bill.

Most side projects stay side projects, and that is fine. You do not need a forecasting model. You need a number you can defend when the next invoice arrives.