Anthropic cancels the Sonnet 5 price hike: $2/$10 stands

We told you twice to budget for it. Claude Sonnet 5 pricing was scheduled to step up from $2/$10 per million tokens to $3/$15 on September 1 — a flat 50% raise on a model a lot of teams quietly made their default. On August 10, Anthropic cancelled it. The introductory rate is now the standard rate. If you resized an automation's ROI at $3/$15, resize it back.

What actually happened

Anthropic's platform release notes for August 10 carry one line: the introductory pricing for Claude Sonnet 5 ($2 / $10 per MTok) is now the standard price, and the previously scheduled increase to $3 / $15 on September 1, 2026 will not occur. The pricing page repeats it as a note on the rate table.

The numbers as they now stand:

• Sonnet 5: $2 input / $10 output per million tokens. Batch is $1/$5.
• Sonnet 4.6: $3 / $15. The previous generation is now the more expensive one.
• Haiku 4.5: $1 / $5. Still the floor if your workload tolerates it.

One detail did not change, and it's the one that moves your invoice. Anthropic's own docs note that Claude 4.7 and later models use a newer tokenizer that produces roughly 30% more tokens for the same text — Sonnet 4.6 and earlier use the previous one. So Sonnet 5 is a third cheaper per token than Sonnet 4.6 while counting more tokens for identical input. Cheaper, yes. Not 33% cheaper.

Why it matters for your business

The good news is boring: your September forecast holds. The useful news is what this says about the shape of the vendor relationship.

A price you build a business case on moved 50% on a scheduled date, then didn't, and both changes arrived as a line in a changelog. Nobody emailed you. There was no contract amendment. The rate card is a web page your vendor edits.

That's not a complaint — it's a design constraint. If a model's price is a variable your vendor controls unilaterally, then the model has to be a variable your code controls too. Concretely, in every system we ship:

Model IDs live in config, not in code. Swapping Sonnet 5 for Haiku 4.5 on a classification step should be an environment variable, not a pull request.

Cost is measured per unit of work, not per month. "$0.004 per processed invoice" survives a price change with an obvious answer. "About $600 a month" does not.

The cheap tier gets tested before you need it. The time to find out Haiku handles your ticket triage is now, while the decision is optional.

We rebuild AI cost models for clients more often than we rebuild the AI. That's the tell. The model layer is the fastest-moving line item in your stack, and the only real defense is a boundary you own — so a vendor's changelog is a config change on your side, not a rewrite.

Sources: Anthropic platform release notes, Anthropic pricing documentation.