TL;DR (July 2026): Claude Sonnet 5 launched June 30 at introductory pricing: $2 per million input tokens, $10 per million output - through August 31, 2026. On September 1 the standard rate takes over: $3 / $15. That is a +50% flip on every Sonnet token, and because Sonnet 5 is the new default for Free and Pro (and the recommended everyday-agent tier), it is the single most predictable cost increase on most Claude bills this year. You have six weeks. The math below shows what the flip does to a real invoice, when Haiku 4.5 becomes the cheaper answer, and the four things worth doing in August rather than discovering in September's invoice.
Intro pricing is a loan against your future bill, and this one has a published due date - which makes it the rare AI cost event you can plan for perfectly. Anthropic set Sonnet 5's launch rate below the 4.6-era rate it replaced ($3/$15), let it become the default everywhere, and dated the reversion. Nothing here is a surprise; the only question is whether your September budget already contains it.
The flip, on a real bill
| Monthly Sonnet volume (in / out) | August (intro $2/$10) | September (standard $3/$15) | Delta |
|---|---|---|---|
| Solo dev - 20M / 6M | $100 | $150 | +$50 |
| Small team - 150M / 45M | $750 | $1,125 | +$375 |
| Product with agent features - 1B / 300M | $5,000 | $7,500 | +$2,500/mo |
The percentage is fixed at +50%; only your volume decides whether that is lunch money or a headcount. If Sonnet is your routing default - and after the Fable 5 metering switch it should be - your Sonnet share is probably rising at exactly the moment its price is about to.
Does the flip change the routing math?
The standard Claude routing stack after September 1 prices like this: Haiku 4.5 at $1/$5, Sonnet 5 at $3/$15, Opus 4.8 at $5/$25, Fable 5 at $10/$50. Two threshold shifts worth noting:
- Haiku's discount vs Sonnet doubles back to 3x (it was 2x during the intro window). Classification, extraction, summarization, and high-volume background jobs that drifted onto "cheap Sonnet" in July belong back on Haiku in September. If a task's quality bar was met by Sonnet-at-intro out of convenience, re-test it on Haiku before the flip does the arithmetic for you.
- Sonnet-to-Opus narrows to 1.67x ($3/$15 vs $5/$25). For hard-but-bounded work where Opus meaningfully outperforms, the premium for stepping up shrinks - the flip makes Opus relatively cheaper, which is worth remembering when you re-tune agent escalation rules.
The two levers that blunt the increase
- Batch processing stays 50% off across models. Anything that tolerates async - nightly jobs, backfills, evals, report generation - priced at batch rates on standard pricing ($1.50/$7.50 effective for Sonnet) still lands below the synchronous intro rate you are paying today. Moving one big workload to batch can absorb the entire flip.
- Prompt caching cuts cached input ~90%. A repeated 100K-token system prompt is the classic case: after caching, its per-call input cost is a tenth. Teams that never enabled caching headers are, in effect, choosing to pay the September increase twice.
The August checklist
- Re-forecast September now. Multiply last month's Sonnet line by 1.5 and put that number in front of whoever owns the budget - in July, not on the October invoice review.
- Re-quote anything client-facing. If you rebill AI usage or quoted fixed prices assuming intro rates, those quotes expire August 31 whether you noticed or not. (Our rebilling guide covers the pass-through mechanics.)
- Re-run the Haiku eval on your three highest-volume Sonnet tasks. A task that passes on Haiku is a permanent 3x saving from September 1.
- Set a burn alert on the Sonnet meter for the first week of September at August's daily average - it will fire at +50% if you changed nothing, which is exactly the confirmation you want, and silence means your mitigations worked.
Full current rate card, limits, and the guardrail patterns are in the Claude API pricing guide - updated for the Claude 5 family. And if your bill's problem is not Sonnet but the frontier tier, start with the July 20 Fable switch instead: that one is not waiting until September.