Claude Opus 5.5: Cheaper, faster, one button short

Hello, dear readers. Do you know what happened on the AI front in the last weeks of September? Of course you do: half the news sites ran the same headline. This month, though, I'm going to take apart what's written under that headline, piece by piece. Our subject is Claude Opus 5.5, which Anthropic released on 22 September 2026.

Let me say this up front: nobody ran this model for the magazine, nobody measured it, nobody did any work on it. This piece is based on Anthropic's announcement page, its price table and the press coverage. Every performance figure you'll see below is Anthropic's own. We couldn't find any independent measurements. If we had, they'd be here.

So what is this Opus 5.5?

Opus is Anthropic's flagship class, and 5.5 is the first member of this family. Opus 5 only came out on 24 July, so the new version arrived less than two months later. We used to wait a year for a new version of a program. Now every two months they tell us to "forget the last one". Oh well, we'll get used to it.

The model is available in the Claude apps and in the Claude API (Application Programming Interface). Its API name is claude-opus-5-5. You can also reach it through Amazon Web Services, Google Cloud and Microsoft Azure. It doesn't matter which cloud you've pledged yourself to, it's on all three (they've left us nothing to fight about, shame on them!). For anyone who doesn't want their data kept, there's a "zero data retention" option too.

The context window, meaning how much text the model can hold in its head at once, gets no number on the announcement page. And I'm not going to make one up.

Now, the money

The real news with this product is the price. Here are the per-million-token prices from Anthropic's page:

| | Opus 5.5 | Opus 5 | |---|---|---| | Input | $4 | $5 | | Output | $20 | $25 | | Cache read | $0.20 | $0.50 | | Cache write | $5 | $6.25 |

Get your calculator out. Input, output and cache writes are 20 percent cheaper. Cache reads are 60 percent cheaper, which is quite a cut. Anyone who feeds the model the same long document or the same codebase over and over gains the most here.

Anthropic also says it's "40 percent cheaper than Opus 5 on typical workloads, with default settings". Fine, but where does that 40 percent come from? If the sticker price dropped 20 percent, the rest has to come from caching, from the model finishing the same job with fewer tokens, or from somewhere else. Anthropic doesn't explain it, and I don't know either. There's another snag: Stokastic wrote that Opus 5 never had published token prices, so the 40 percent has nothing to be compared against. Yet Anthropic's own page shows the prices above for Opus 5. The two don't add up. My advice: trust the sticker prices in the table, and keep the 40 percent in your back pocket as the company's own claim.

Then there's "fast mode". Input is $8 and output $40 per million tokens, exactly double. In return, Anthropic says it's "up to 2.5 times faster". I'm underlining "up to", because that's a ceiling, not an average. It's available in Claude Code and on the Claude Platform. If you've got a customer waiting, or a build loop that eats your minutes, the money might make sense. On a batch job that runs by itself overnight, it's money down the drain.

In normal mode, too, Anthropic says output is more than 30 percent faster than Opus 5. TechCrunch puts this speed down to less compute being spent serving the model. That's probably where the price cut comes from as well: the machine has less work to do.

What do the numbers say?

Anthropic's claim is a bold one: Opus 5.5 supposedly works "at the level of Claude Fable 5.1 on most tasks". Fable is the company's bigger, pricier model. A cheap model keeping up with the expensive one is always good news, but let's look at the table first (reminder again: it's all Anthropic's):

| Benchmark | Opus 5.5 | Fable 5.1 | Opus 5 | |---|---|---|---| | Terminal-Bench 4.0 | 66.4% | 55.8% | 52.3% | | FrontierCode v1.1 | 54.4% | 50.3% | 48.0% | | CursorBench 4.0 | 57.8% | 51.8% | 46.6% | | GDPval-AA v2.1 (Elo) | 1846 | 1735 | 1708 | | Humanity's Last Exam (with tools) | 67.7% | 65.6% | 63.6% | | OSWorld 2.1 (partial) | 81.8% | 80.7% | 74.0% |

The first thing that caught my eye was Terminal-Bench. This benchmark checks whether the model can get work done on its own at a command line, and the gap over Opus 5 is 14 points. There's no bigger jump in the table. CursorBench shows an 11-point gap too. If you're after a model that writes code, gets around a terminal and finishes jobs inside a software project, these two rows are the strongest evidence on paper.

As for the "Fable level" claim, going by the table it's almost too modest, because Opus 5.5 is ahead of Fable 5.1 on all six of these benchmarks. But careful! The phrase "on most tasks" is the company's own. So there are jobs where Fable is still better, and they don't show up in this table.

And against the competition? Anthropic's own page lists nine benchmarks, and OpenAI's GPT-6 Astra leads on two of them: 41.4 percent against 40.0 on AutomationBench, and 64.6 against 58.7 on Terminal-Bench-Science. The first is close to a tie. The second is a really big gap, and if you're someone who wrestles with scientific computing in a terminal, you should care about it. I'll admit I appreciated the company showing the rows where it loses in its own table. You don't see that much in a brochure.

Changes in behaviour

This is where a button has gone missing. In Opus 5.5, "thinking" (the stage where the model reasons to itself before answering) can no longer be switched off. What does that mean? Even when you want a short answer to a simple question, the model will think first. Thinking means tokens, and tokens mean money. People who switched thinking off for a quick, cheap answer lose that convenience. And don't forget that Anthropic's 40 percent figure was worked out "with default settings". The default is now the only setting. I think the reasoning behind this decision should have been spelled out in the announcement. It isn't.

The other change is a nicer one. According to TechCrunch, the model uses less jargon and puts the important information first. Anyone who has dealt with an assistant that says "oh, by the way, here's what actually mattered" at the end of a long answer will know what that's worth.

There's also a safeguard called "preserved thinking". Its purpose is to block distillation, that is, to stop someone from collecting this model's outputs and training their own model on them. It applies to API accounts opened on or after 31 August 2026. I couldn't work out from the documentation exactly what the safeguard does technically, so I can't tell you more on that. But if you're a developer about to open a new account, check the documentation to see whether it affects your work. You MUST.

The safety side

On cybersecurity and biology, Opus 5.5 comes with protections similar to Fable 5.1's. Some requests are handed off to another model, and Anthropic says this is done openly. For organisations doing legitimate work in these areas, a verification programme called the Life Sciences Verification Program has been opened, and the existing Cyber Verification Program is being expanded. So if you're a lab or a security firm, you can show your ID and get through the door.

Before release, the model was also tested by outside organisations such as METR and Frontier Design. TechCrunch also points out that this is Anthropic's first model since Dario Amodei came out in support of progress on frontier models keeping pace with alignment work. How to read a two-month release cycle alongside "we're being careful" is a topic for another piece (and I don't have enough information to write that piece).

A price war?

Some headlines billed this release as a "price war". I couldn't find a source showing competitors answering with their own prices. The price did drop, that's true. The war part is, for now, a product of the headline writers' imagination.

Anthropic also announced that Sonnet 5.5 and Haiku 5.5 will arrive "in the coming weeks". We didn't check for this piece whether they've come out. If they do, we'll look at them separately.

Verdict

On paper, Opus 5.5 is a pretty good bargain: cheaper, faster, and by the company's own numbers it beats the model a class above it on plenty of tasks. The jump in code and terminal work is the most convincing part of the table. If you ask me, anyone using Opus 5 for coding has little reason not to switch. But not being able to turn thinking off is a real loss for people used to short, cheap answers. And the 40 percent savings claim is just a claim until someone independent measures it.

One more thing, because I say it with every new release: if you're going to upgrade a working system by swapping the model, run the old one and the new one side by side on your own test set BEFORE the switch. Somebody else's table doesn't measure your work.

If independent measurements come out, or if something interesting happens to you with this model in your own work, write in. I can't quite bring myself to call them letters anymore, but you know what I mean. See you next month, goodbye!