I Told You Claude's API Got 50% More Expensive. Anthropic's Pricing Page Says the Increase Never Happened.
On September 18 I published a post arguing that GitHub cut Copilot credits and Anthropic raised Claude's API price in the same week. The Anthropic half of that is wrong. Anthropic's pricing documentation says the Claude Sonnet 5 rate of $2 per million input tokens and $10 per million output tokens, originally introductory through August 31, "is now the standard price," and that the scheduled increase to $3/$15 on September 1 "will not occur."
There was no 50% increase. I reported one, and the correction is the post.
What I got wrong and how
The mechanism matters more than the mistake. That post cited a single blog analysing the September 1 pricing changes. It was detailed, its numbers were internally consistent, and a second source described the same changes in general terms. My own fact-check log for that piece says so explicitly: the figures came from one detailed source, corroborated in general terms by a second, and were "treated as verified."
That is not verification. Anthropic publishes a pricing page. When the claim is "this vendor changed this price," the vendor's own pricing documentation is the primary source, and a blog post about the pricing change is a secondary account of it. I had a rule for this and did not apply it, because the secondary source was specific and confident, and specificity reads like accuracy.
The most likely explanation is straightforward: a price increase was scheduled, publicised, written about in advance, and then cancelled. Secondary coverage of a scheduled change ages badly the moment the schedule changes, and nothing about the article that reported it looks different afterwards.
I am not correcting the GitHub Copilot half of that post. I have not re-verified it, so I am not going to assert anything about it here either way.
Opus 5.5 moves the other direction
Anthropic released Claude Opus 5.5 on September 22. The pricing, from the same documentation I should have read last week:
| Opus 5 | Opus 5.5 | |
|---|---|---|
| Input | $5 / MTok | $4 / MTok |
| Output | $25 / MTok | $20 / MTok |
| Cache read | $0.50 / MTok | $0.20 / MTok |
| 5m cache write | $6.25 / MTok | $5 / MTok |
| 1h cache write | $10 / MTok | $8 / MTok |
| Batch in/out | $2.50 / $12.50 | $2 / $10 |
Base pricing is down 20% on both legs. Anthropic says the model performs at Claude Fable 5.1's level on most work and costs about 40% less to run than Opus 5 on typical workloads.
The cache read line is the story
Look at that cache read row again. It went from $0.50 to $0.20, which is 60%, three times the headline cut.
That is not a rounding artifact of the 20% reduction. Every Claude model prices cache hits at 0.1x its base input price. Opus 5.5 uses 0.05x, a multiplier otherwise reserved for Fable 5.1 and Mythos 5.1. So Opus 5.5 got a price cut and a structural change to how cache reads are billed, and only one of those made the headlines.
For a chat app sending short, varied prompts, the 20% base cut is the number that matters. For a long-running agent, it is not, because that workload is mostly cache reads. An agent with a large stable system prompt, a tool schema, and a growing conversation history re-sends that context every turn, and after the first write it is all cache hits.
Take a coding agent holding roughly 100k tokens of cached context across 40 turns in a session. On Opus 5, four million cached input tokens at $0.50 is about $2.00. On Opus 5.5, at $0.20, it is $0.80. Add the 20% cut on the uncached input and the output and the session total falls well past 20%, because the biggest line item fell 60%.
If you run agents on Opus, re-run your own numbers with the split between cache reads and fresh input. The saving is larger than the announcement implies, and it is larger in proportion to how much context you hold.
Two smaller things from the same page. Fast mode on Opus 5.5 is $8/$40 against $10/$50 on Opus 5. And the tool use system prompt is 286 tokens for auto and none on both models, but Opus 5 charges 406 tokens for any and tool while Opus 5.5 lists a single 286 figure.
What I'd actually do
Change where you read pricing from. For any vendor whose API you pay for, the pricing page is the source and everything else is commentary. That sounds obvious written down, which is exactly why it is worth writing down, because I skipped it while running a pipeline that has a fact-check step in it.
The pipeline change I am making: any claim about a vendor's pricing has to cite that vendor's own pricing documentation, not an article about it. Secondary sources can establish that something happened. They cannot establish a number that the vendor publishes themselves.
If you are running Opus 5 in production, the migration is worth doing on price alone, and more so if your workload is cache-heavy. Check the model string change to claude-opus-5-5 and measure your own cache read ratio before and after rather than trusting my arithmetic.
The part I am less sure about
I do not know precisely when Anthropic cancelled the Sonnet 5 increase, or whether it was ever in effect for any window. The pricing page states the current position, not the history. It is possible the increase briefly applied and was reversed, in which case the September 18 post was less wrong than it looks, though still sourced badly. I would rather over-correct here than defend a number I cannot support.
I am also aware that a post correcting my own reporting is a comfortable genre. It signals rigour at low cost. The test is not this paragraph, it is whether the next pricing claim on this site cites a pricing page. Hold me to that one instead.
Author
Lukas
@lukcombinator