Claude Opus 5.5 Price: Cheaper Than Anthropic Says It Is

The Quick Answer

Anthropic launched Claude Opus 5.5 on Tuesday promising it costs 40 percent less to run, and the company is understating its own discount. That 40 percent measures the new model against Opus 5, the model it replaces. But Anthropic also says Opus 5.5 does the work of Claude Fable 5.1, the flagship it shipped three weeks earlier — and against Fable 5.1's posted prices, the cut is 60 percent. Anyone paying today for that level of work is the buyer the 40 percent does not describe.

Anthropic Measured Against The Older Model

One sentence in the announcement carries two different comparisons. Opus 5.5 "performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5." The quality claim points at Fable 5.1. The price claim points at Opus 5. Those are different models with different price cards.

Comparing a new model to the one it replaces is the conventional thing to do, and Opus 5 is the predecessor in the same line. It also happens to understate what the relevant customer saves. Someone paying Fable 5.1 rates today is doing it because they need that standard of work, and Opus 5.5 is the first cheaper model Anthropic has claimed reaches it. Their saving is on the rate card in plain sight: $4 against $10 going in, $20 against $50 coming out.

Anthropic published the smaller of the two numbers.

The Numbers

Anthropic's posted rates, per million tokens, as of Tuesday, Sept. 22, 2026.

Anthropic never posted text-in and text-out rates for Opus 5, which is why its own 40 percent claim cannot be checked against a price list.

What "$10 Per Million Tokens" Actually Means

These are what a developer pays to call the model from their own software, metered by how much text moves in each direction.

Tokens are the unit the meter counts — pieces of words rather than whole ones. "Per million tokens" is the mark where the charge lands. Claude Fable 5.1 bills $10 every time a million units of text goes into the model, and $50 every time a million comes back out. Output costs five times input on every model here, because generating text is the expensive half.

So a job that sends a long document in and gets a short answer back is billed mostly at the input rate. A job that sends a short instruction and gets pages back is billed mostly at the output rate. Both of those halves drop by 60 percent moving from Fable 5.1 to Opus 5.5.

Where The 60 Percent Shrinks

Sixty percent covers text in and text out. It does not cover the third line, and for some buyers that line is the whole bill.

When a model re-reads text it has already been given — the same large codebase, the same long brief, hundreds of times a day — it is billed at a separate, much lower rate. Anthropic says that line makes up the majority of cost in agentic and coding work. Opus 5.5 charges $0.20 per million there. Fable 5.1 charges $0.25. That gap is 20 percent, not 60.

So the real saving depends on the shape of the job. Work dominated by fresh text in and fresh text out saves close to 60 percent. Work dominated by re-reading the same context saves closer to 20. Anthropic's 40 percent sits between those two, which is what a blended figure looks like — it is just blended against a different model than the one the quality claim names.

The Claim It All Rests On

Every number above depends on "performs at the level of Claude Fable 5.1 on most work." If that is not true, a Fable 5.1 customer is not looking at a 60 percent discount, they are looking at a cheaper model that does less.

"Performs at the level of" and "on most work" are both carrying weight, and neither phrase names the work that falls outside "most." Anthropic did attach numbers on Sept. 22, 2026:

What is missing is the comparison column. The same tests run on Claude Fable 5.1 and Claude Opus 5 are absent, as is any independent run. A reader can see every score and still not have the one thing that would confirm the sentence the pricing argument rests on.

The Market That Priced The Release

Polymarket US, which bills itself as the first federally regulated prediction market in the United States, ran a market on exactly this event: "Next Claude Opus Released?", filed under Tech and Model Releases, with two rungs, by September 30, 2026 and by October 31, 2026. The chart on it opens September 11.

The day before the release the September rung was not a lock. It traded down to about 70 percent in the evening, climbed back through the night, sat in the mid-90s through Tuesday morning, and went to 100 percent around midday, when the model shipped.

Both rungs are settled by the release itself, eight days before the September date. A penny for No is what a market charges when the answer is already known.

What that board could not tell anyone is the part that turned out to matter. It asked whether the next Opus would arrive, not what it would cost.

What settles the pricing question is the next few weeks. The rates are already public. Whether any competitor answers with a cheaper tier of its own is the thing to watch. It becomes a market event when a rival has to match it.

About The Numbers

A price per million tokens is what Anthropic charges developers for text going into a model and text coming back out; it is not a consumer subscription rate. Anthropic has not posted text-in and text-out rates for Opus 5 alongside the new card, so the 40 percent run-cost claim cannot be checked against a sticker price here — the two Opus 5 figures that are public are in the table. The 60 percent figure is arithmetic on posted rates: $4 against $10, and $20 against $50. The 20 percent figure is the same arithmetic on the re-reading line, $0.20 against $0.25. On a prediction market the price is the chance, so a No side quoted at 1 cent puts the Yes answer at about 99 percent. Release-market prices and chart readings are from Polymarket US on the afternoon of Tuesday, Sept. 22, 2026. The benchmark figures are single published scores from Anthropic's own testing, each on one named test, and the 81.8% on OSWorld 2.0 is marked partial. The 1846 Elo is a rating on a comparison scale, so it carries meaning only against other models' ratings on that same scale, which are absent here. Trading volume on the release market is left out, since the venue's own page reported none for either rung.