Claude Sonnet 5 Price Increase Cancelled: $2/$10 Stays
The Short Answer
The increase is off. Claude Sonnet 5 stays at $2 per million input tokens and $10 per million output tokens — permanently.
Anthropic’s official pricing documentation now carries this note:
The $2/$10 per million input/output token pricing for Claude Sonnet 5, announced at launch as introductory pricing through August 31, 2026, is now the standard price. The previously scheduled increase to $3/$15 per million input/output tokens on September 1, 2026 will not occur.
If you read anywhere that you have until August 31 to lock in cheap Sonnet 5, or that your bill rises 50% next month, that information is stale. A large number of articles published in early August 2026 still describe the increase as upcoming.
What The Price Actually Is
| Token type | Claude Sonnet 5 |
|---|---|
| Base input | $2 / MTok |
| Output | $10 / MTok |
| Cache hits & refreshes | $0.20 / MTok |
| 5-minute cache write | $2.50 / MTok |
| 1-hour cache write | $4 / MTok |
A representative 30K input / 5K output task costs about $0.11.
For context in the current Claude lineup: Opus 5 is $5/$25, Haiku 4.5 is $1/$5, Fable 5 and Mythos 5 are $10/$50, and the older Sonnet 4.6 remains at $3/$15 — meaning Sonnet 5 is now permanently cheaper than the model it replaced.
The Tokenizer Caveat That Still Applies
Cancelling the price rise does not mean your bill matches a naive comparison against older models. 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 tokenizer.
So Sonnet 5 at $2 versus Sonnet 4.6 at $3 is not a flat 33% saving. Adjusting for the tokenizer, the real-world input saving is closer to 13%, and output economics depend on how verbose the model is on your specific workload. Still a saving — just not the one the headline rates imply.
Two Other Pricing Modifiers
Data residency. For Claude 4.6 and later, requesting US-only inference via inference_geo: "us" applies a 1.1x multiplier on every token category — input, output, cache writes and cache reads. Global routing is the default and carries no premium. The same 1.1x applies to Microsoft Foundry deployments using the US Data Zone Standard type.
Regional endpoints. On Bedrock and Google Cloud, regional and multi-region endpoints carry a 10% premium over global endpoints for Sonnet 4.5, Haiku 4.5, Opus 4.5 and all later models.
If you assumed $2/$10 and you’re pinned to a US region, your effective rate is $2.20/$11.
Why This Matters Beyond Anthropic
August 2026 has run in two directions at once:
| Date | Change | Direction |
|---|---|---|
| Jul 30, 2026 | GPT-5.6 Luna cut 80% → $0.20/$1.20; Terra cut 20% → $2/$12 | ⬇ |
| Aug 10, 2026 | Claude Sonnet 5 increase cancelled; $2/$10 permanent | ⬇ |
| Aug 13, 2026 | Gemini 3.7 Flash launches at $0.75/$3.75 intro | ⬇ |
| Aug 16, 2026 | DeepSeek V4 raises prices up to ~1,100%, adds peak rates | ⬆ |
The pattern is not “AI is getting more expensive” or “AI is getting cheaper.” It is divergence. Well-capitalised labs with their own capacity are competing on price at the mid and low tiers. Capacity-constrained providers are rationing with price. DeepSeek’s increase and Anthropic’s cancellation happened six days apart and point in opposite directions for the same underlying reason: who has compute.
Anthropic reported Q2 2026 revenue of $11.5 billion, up 143% from Q1. A company on that trajectory does not need a 50% mid-tier price rise, and taking one while OpenAI cuts Luna by 80% would have been a strategic gift.
What To Do
- Cancel any migration you started to beat the September 1 deadline. The reason no longer exists.
- Correct your cost models. If your 2026 H2 forecast assumes $3/$15 from September, you are over-budgeted by 50% on Sonnet 5 line items.
- Check
inference_geo. If you set it to"us"for compliance, your real rate is 1.1x. Confirm you actually need it. - Recheck the tokenizer assumption if you migrated from Sonnet 4.6 and your bill did not fall as much as expected. That’s the ~30% tokenizer difference, not a billing error.
- Treat vendor pricing pages as the only source of truth. This story is a clean case study: the secondary coverage was wrong for a week after the primary source changed.
Last verified: August 17, 2026, against Anthropic’s official pricing documentation.