Anthropic shipped Claude Opus 5.5 on Tuesday, September 22, at $4 per million input tokens and $20 per million output tokens. That is 20 percent below Opus 5 on both lines, and cheaper still against the $10/$50 Anthropic charges for Fable 5.1. On the benchmarks the company published for both models, it beats its predecessor everywhere — 81.8 percent on the partial OSWorld 2.0 agentic tasks against Opus 5’s 74.0 — and roughly matches the premium tier on most work. Hacker News handed it 1,686 points and 1,025 comments, which is the modern equivalent of a ticker-tape parade.

The consensus on the release writes itself: the AI price war is here, capability is commoditizing, margins are collapsing, buyers win. Fine. The consensus is also missing the interesting part, which is that the headline number is a rate, and a rate means nothing without the unit of account — and the seller defines the unit.

The 40% Doesn’t Come From the Price Sheet

Start with the arithmetic Anthropic asks you to accept. The rate card fell 20 percent, not 40. The 40 percent figure — “cheaper to run than Opus 5,” in the company’s own framing — is a blended claim. Cache reads did drop hard, from 50 cents to 20 cents per million. What closes the rest of the gap is not on the pricing page.

It is on the effort page. Per Anthropic’s September effort documentation, Opus 5.5 defaults to medium. Opus 5 and every Opus before it defaulted to high. An application that changes nothing but the model string — the one-line migration the launch is quietly selling — now receives a different quantity of reasoning with the same prompt and the same parameters.

Multiply that across every agentic loop in production and the arithmetic stops being tidy. The bill is rate times tokens, and tokens per completed task is a number no vendor publishes.

A staff engineer at a mid-sized insurance-software vendor, typing in the Slack channel where his team ran the swap on Tuesday afternoon, put it about as well as procurement ever will: “We changed one string. Our evals moved because the model stopped thinking as hard, the invoice went down, and the ticket count went up. Which of those is the improvement?”

To be fair to Anthropic: it documents the behavior, exposes an effort control, and sells a fast mode at double the standard rate that returns output up to 2.5 times faster. That is transparency, by the standards of this industry. It is also the tell. Three dials now sit on the product — how hard the model thinks, how fast it answers, what it costs — and the company owns all three.

Fable 5.1’s Customers Just Got Marked Down

Here is the part enterprise buyers should sit with. Fable 5.1 sits at $10 and $50 per million tokens, double Opus 5.5 across the board, and Anthropic’s launch copy has a model at less than half the rate performing “at the level of” it on most work. That is not a rival undercutting the premium tier. That is the premium tier being repriced by the house, on a Tuesday, with no mechanism for anyone who sized this quarter’s budget around it.

The same logic now applies to speed. Latency used to be a property of a system you engineered for. As of this week it is a SKU with a multiplier attached.

I don’t blame Anthropic. Any lab would do this, and the rest will. Capability gains at the frontier are measured in single digits on the vendors’ own charts — 89.0 against 88.4 across two Anthropic models is noise with a marketing department. When you can no longer sell a step change in intelligence, you sell the terms. But a buyer reading the rate card as a price signal is reading a number the seller sets, denominates, and schedules for depreciation. Nobody else in American commerce gets that arrangement. A grocer cannot define the ounce.

Ask for Cost Per Completed Task

The migration guidance shipped alongside the release contains the whole argument in one sentence: compare behavior, because a model-string change can alter reasoning even when prompts and parameters don’t. So compare cost per completed task. Tokens per resolved ticket, per merged pull request, per closed reconciliation — the same discipline that made shelf unit-pricing and the APR non-negotiable.

That is the uncomfortable part for anyone who spent the last several years insisting this market needs nothing but room to compete. The competition is working. Capability up, price down, same release. But the signal that is supposed to discipline the seller is issued by the seller, in units the seller chooses, on a tier the seller can retire the moment a cheaper sibling catches up. Price discovery needs a common measure. Right now the measure is a marketing decision.

every.to’s reviewer described the upgrade as personality rehab as much as an intelligence jump — Opus 5, in that publication’s words, “had a bit of a personality problem,” argued with change requests, ran long on simple tasks. A model that argues less and, by default, thinks less is not a smaller story than a benchmark number. It’s the product.

The 40 percent is real. It just isn’t the number you think it is.

Sources