If you budgeted for Claude Sonnet 5's API cost to jump on September 1, 2026, you can stand down. Anthropic confirmed on August 10 that the $2 per million input tokens / $10 per million output tokens rate, originally framed as introductory pricing, is now Sonnet 5's permanent, standard price. The previously scheduled increase to $3/$15 per million tokens will not happen.
This matters because that increase was real. It sat on Anthropic's own pricing page for weeks, and a lot of teams planned around it. If you're still seeing "Sonnet 5 jumps to $3/$15 on September 1" anywhere, including in some cost calculators and comparison posts, that information is now stale. Anthropic's current pricing documentation and its Sonnet 5 launch post both confirm the reversal.
01What Changed, Specifically
| Rate Type | Scheduled (Through Aug 31) | Scheduled (From Sep 1) | Confirmed Permanent |
|---|---|---|---|
| Input Tokens | $2 / MTok | $3 / MTok | $2 / MTok |
| Output Tokens | $10 / MTok | $15 / MTok | $10 / MTok |
That 50% rate increase simply isn't happening. Anthropic's official framing is that the "introductory" label no longer applies. This is now the standard rate, full stop.
See what this means for your actual workload
Run your token volumes against the confirmed permanent rate02Why This Is Easy to Miss Right Now
Sonnet 5 launched with a pricing structure that had a built-in deadline: cheap through August 31, standard (and pricier) from September 1. That deadline was repeated across technical documentation, developer blogs, and third-party trackers between late June and early August.
Because the cancellation is fresh, plenty of legacy content hasn't caught up. If you're comparing model costs by reading an older post rather than checking a live source, there's a real chance you're working off the outdated $3/$15 numbers.
03The Part That Doesn't Show Up in the Headline: The Tokenizer
Here's the honest caveat, and it's the one worth actually budgeting around. Sonnet 5 and later models use an updated tokenizer.
Per Anthropic's documentation, this tokenizer produces roughly 30% more tokens for the same text compared to the tokenizer used by Sonnet 4.6 and earlier. The exact increase depends on your workload shape, with code, structured JSON data, and non-English text typically affected the most.
Key takeaway"The price per token didn't go up" isn't the same as "your bill won't go up." If you're migrating from Sonnet 4.6 to Sonnet 5, identical prompts can yield higher billable token counts even at the lower per-token rate. Running your actual prompt logs through the Sonnet 5 calculator is the fastest way to see the real number instead of estimating it.
04Prompt Caching and Batch Discounts Still Apply
Nothing about the caching or batch discount structure changed — only the base rate stopped moving. At the confirmed $2/$10 base rate:
- Cache writes (5 min)1.25x base — $2.50/MTok
- Cache writes (1 hour)2x base — $4.00/MTok
- Cache hits (reads)0.1x base — $0.20/MTok (90% off)
- Batch API, input50% off — $1.00/MTok
- Batch API, output50% off — $5.00/MTok
These stack with each other, meaning a cache-heavy, batch-eligible workload lands well under the headline sticker rate.
05What to Do With Your API Budget
If you reserved budget for the September 1 step-up, you can safely reallocate those funds. If you migrated from Sonnet 4.6 and assumed a flat cost comparison, the tokenizer difference above is the thing to check instead of the rate. Either way, the fix is the same: re-run your actual prompt templates through Sonnet 5 and measure real token counts rather than carrying over estimates from an older model.
Model your real Sonnet 5 costs
Caching, batching, and multi-turn conversations included