The one-line version: Claude Sonnet 5's launch discount ends. Per Anthropic's pricing page, the introductory rate of $2 / $10 per million input/output tokens runs through August 31, 2026; on September 1 the model reverts to standard $3 / $15 — a flat 50% increase. If you run an agent on Sonnet 5, your token bill goes up by half on that date for identical work. This is the number to plan around, and the good news is you have the whole month to decide what to do about it.
What actually changes#
Everything moves by the same 1.5x. This is the part most "prices went up" write-ups miss: the increase isn't just on the headline base rate — it lands on every pricing line at once.
| Line item | Intro (through Aug 31) | Standard (from Sep 1) | Change |
|---|---|---|---|
| Base input | $2 / MTok | $3 / MTok | +50% |
| Output | $10 / MTok | $15 / MTok | +50% |
| Cache read (hit) | $0.20 / MTok | $0.30 / MTok | +50% |
| 5-min cache write | $2.50 / MTok | $3.75 / MTok | +50% |
| 1-hour cache write | $4 / MTok | $6 / MTok | +50% |
| Batch input | $1 / MTok | $1.50 / MTok | +50% |
| Batch output | $5 / MTok | $7.50 / MTok | +50% |
That uniformity has a direct consequence for how you respond.
Why caching and batching won't save you#
The instinct when a price rises is to reach for the discount levers — turn on prompt caching, move non-urgent work to the Batch API. Those are good ideas anyway, but they do not blunt this specific increase, because the increase applies to the discounted rates too.
A cache hit is still one-tenth of base input — but base input went from $2 to $3, so the cache read went from $0.20 to $0.30. The Batch API still takes 50% off — but off a base that's now 50% higher. Your caching ROI is unchanged; your absolute bill is multiplied by 1.5 regardless of how aggressively you cache. If you were already caching and batching, you've already banked those savings, and they can't be spent twice to offset September 1.
The only two things that beat a uniform per-token increase are sending fewer tokens or changing which model processes them. Discounts that scale with the base rate ride the increase up with it.
The bill, worked#
Take a mid-size agent — a support or coding assistant running steadily — at 500M input tokens and 25M output tokens a month (agents are input-heavy: system prompts, tool results, and history dominate).
Today (intro): 500M × $2/M = $1,000 input, plus 25M × $10/M = $250 output → $1,250/month.
September 1 (standard): 500M × $3/M = $1,500 input, plus 25M × $15/M = $375 output → $1,875/month.
That's +$625/month — $7,500/year — for byte-for-byte identical work. To size your own exposure, pull last month's Sonnet 5 spend from the Claude Console and multiply by 1.5. That's your September run-rate before you change anything.
The subtlety that costs more than the headline#
Here's the trap. On September 1, Sonnet 5 will cost exactly what Sonnet 4.6 costs — both at $3 / $15. So why not treat them as interchangeable and pick on quality alone?
Because they don't tokenize the same text the same way. Anthropic's pricing note is explicit: models from Claude 4.7 onward use a newer tokenizer that produces roughly 30% more tokens for the same text, while Sonnet 4.6 and earlier use the previous one. Sonnet 5 is on the new tokenizer; Sonnet 4.6 is not. At equal per-token prices, the model that emits ~30% more tokens for the same prompt-and-response is the more expensive one per task — a cost that never appears on the price sheet. We pulled that thread in full in Sonnet 5's tokenizer tax; September 1 is when it stops being an intro-price footnote and starts being a live line item, because the discount that was papering over it disappears.
The honest caveat: Sonnet 5's stronger reasoning can win the tokens back by finishing a task in fewer turns or with less retrieved context. Whether it does is a property of your workload, not a fact you can read off a table — which is exactly why the move is to measure.
Four levers that actually work#
In rough order of effort-to-payoff:
- Send fewer tokens. The cheapest token is the one you don't send. Trim bloated system prompts, cap tool-result payloads before they re-enter context, and summarize conversation history instead of replaying it. On an input-heavy agent this is the highest-leverage change and it compounds with every other lever.
- Route by difficulty. Most agent turns are easy. Send those to Haiku 4.5 at $1/$5 and reserve Sonnet for the turns that genuinely need it. A difficulty router that offloads even half your traffic can more than absorb the 50% rise on what's left. See when the budget model loses for where cheap tiers break down.
- Re-evaluate Sonnet 4.6 at the now-equal price. With the sticker prices identical from September 1, the lighter tokenizer can make 4.6 cheaper per task for text-heavy work. Benchmark both on your real prompts; if 4.6 holds quality, the token-count edge is free savings.
- Test an open-weight backend for cost-dominant paths. For high-volume, quality-tolerant work, a self-hosted or hosted open model changes the unit economics entirely — the comparison we ran in Kimi K3 vs Claude Sonnet 5 for an agent backend and cheapest tokens vs frontier default.
What to do before August 31#
The deadline is a forcing function, not an emergency. One useful thing to do while intro pricing is still live: run your candidate models — Sonnet 5, Sonnet 4.6, Haiku 4.5, maybe an open-weight backend — through your real evaluation set now, and record both quality and token counts. Then the September 1 decision is a lookup, not a scramble. The increase is real and uniform; the way to spend less through it is to have already measured which model does your specific job for the fewest tokens.
If you decided Sonnet 5 was the cheap way to run agents in June — Anthropic pitched it exactly that way — that thesis isn't wrong, it just gets 50% more expensive on September 1. Re-run the numbers you ran then. The answer may still be Sonnet 5. Now you'll know instead of assume.
