Pin Down the Current Price First
Per Anthropic's official pricing documentation, Claude Sonnet 5 is now $2 per million input tokens and $10 per million output tokens, with $2.50 for 5-minute cache writes, $4 for 1-hour cache writes and $0.20 for cache hits. The Batch API applies a 50% discount, making it $1/$5. A dedicated note in the docs states that the $2/$10 announced at launch as introductory pricing through August 31, 2026 is now the standard price, and that the previously scheduled increase to $3/$15 on September 1, 2026 will not occur. This is worth its own piece because of the timing: today is the date the increase was supposed to take effect, and plenty of third-party pricing guides, comparison tables and cost analyses still list $3/$15. If you built a budget last month by copying one of them, your number is too high. Teams that produced annual estimates should go back and correct them. One comparison worth noting: Sonnet 4.6 and 4.5 remain at $3/$15, so Sonnet 5 is a third cheaper than its predecessor rather than level with it.
The Rate Held, but Your Bill May Still Have Moved
Easier to overlook than the rate is the tokenizer. Another note in the official docs is explicit: Claude 4.7 and later models (plus Mythos Preview) use a newer tokenizer that produces approximately 30% more tokens for the same text, with the exact increase depending on content and workload shape. Sonnet 4.6 and earlier use the previous tokenizer. Put the two together: migrating from Sonnet 4.6 ($3/$15, old tokenizer) to Sonnet 5 ($2/$10, new tokenizer) cuts the rate by a third while raising token counts by roughly 30%, and the net effect varies by workload. You cannot conclude anything from unit price alone. For a migration decision, the reliable method is running a batch of real requests through both and comparing total spend, not comparing per-million rates. Independent testing has also reported that at higher effort settings Sonnet 5 generates many thinking tokens and can cost more per task than Opus 4.8. That comes from third-party testing rather than Anthropic, but it points the same direction: price the task, not the token.
Check Separately If You Go Through a Cloud Provider
If you access Claude through Amazon Bedrock, Google Cloud Vertex AI or Microsoft Foundry, the cloud provider bills you and terms are independent — the official docs direct you to each provider's pricing page. A few other easy traps: for Claude 4.6 and later, pinning inference to the US (setting inference_geo to us) applies a 1.1x multiplier, and regional endpoints on Bedrock and Google Cloud carry a 10% premium over global endpoints.
via: Anthropic's pricing documentation, Anthropic: Introducing Claude Sonnet 5