Short version
Claude Opus 5.5 is Anthropic's new everyday top model. In Anthropic's own words, it "performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5." The price per word fell by a fifth, and the model needs fewer words to finish a job, which is where the rest of the saving comes from. It also writes more clearly, puts the important thing first, and behaves more safely when it works on its own. It is available now in the Claude apps, in Claude Code and to developers.
The race goes on: three top models in three weeks
September has been a month of flagships. On 1 September Anthropic released Claude Fable 5.1, its most capable model, built for work that runs for hours. Two days later OpenAI answered with GPT-6 Astra. Both came in at the same high price. Now Anthropic has moved again, this time not upwards but sideways: a model that gets close to the top for much less money.
Prices for these models are quoted per million tokens. A token is a piece of a word, and Anthropic's rule of thumb is that one token is about three quarters of an English word. "In" is what you send the model, "out" is what it writes back, and writing costs more than reading.
| Model | Maker | Released | Price per million tokens, in / out |
|---|---|---|---|
| Claude Fable 5.1 | Anthropic | 1 September | $10 / $50 |
| GPT-6 Astra | OpenAI | 3 September | $10 / $50 |
| Claude Opus 5.5 | Anthropic | 22 September | $4 / $20 |
Prices checked on the makers' own pages, Anthropic and OpenAI, on 23 September 2026.
Anthropic says the new model now leads on its tests of coding, of operating a computer and of office work. The most telling one is GDPval-AA, a test of real work across 44 occupations: there Anthropic reports that Opus 5.5 at its everyday setting beats GPT-6 Astra at full stretch for about a fifth of the cost per task. These are Anthropic's own measurements, and Anthropic adds its own caution in the same breath: at this level "benchmark margins have become a less reliable guide to real-world differences", and in its own use the gap to Fable 5.1 is "narrower than these scores suggest."
There is a twist in the timing. A week before the launch, Anthropic's CEO Dario Amodei argued that the industry should slow down: "We must slow the pace at which we improve the capabilities of AI models. Progress will still seem fast." Opus 5.5 is Anthropic's first release since that essay, and it shows what he meant. Pacing, in his words, "does not mean halting model training or technical progress". The race goes on. This model competes on cost and care, not on raw power.
What this means for you: the best model is no longer a fixed thing you buy once. In three weeks the list price for work at close to the top level fell from $10 and $50 to $4 and $20, and it will keep moving.
Much cheaper: what 40% less looks like
The saving comes from two places at once. The price per token fell, and the model uses fewer tokens to finish the same job. Anthropic puts the combined effect at "a 40% drop in costs" on typical work compared with Opus 5.
| Per million tokens | Opus 5 | Opus 5.5 | What it means |
|---|---|---|---|
| Reading (input) | $5 | $4 | 20% less for everything you hand it |
| Writing (output) | $25 | $20 | 20% less for everything it writes |
| Re-reading (cache reads) | $0.50 | $0.20 | 60% less, and on long jobs this is most of the bill |
Prices from Anthropic's pricing page, checked on 23 September 2026.
The last row needs a word. On a long job the model reads the same material again and again: your documents, its own notes, the conversation so far. It keeps that material in a short-term memory called a cache and pays a small fee each time it re-reads it. Anthropic says these re-reads "make up the majority of agentic and coding work costs", which is why that 60% cut matters more than the headline 20%.
Anthropic's own examples show what that looks like on a real job. In one test, Opus 5.5 and Opus 5 both analysed a proposed merger, built a financial model in Excel and turned it into a board presentation. They reached the same conclusion, but Opus 5.5 finished in 63 minutes instead of 93 and cost 50% less. In another, it translated a widely used piece of web software into a different programming language in 9.5 hours, against 12 for Fable 5.1, for 51% less money.
Early customers report the same pattern. The trading firm Optiver says Opus 5.5 matched the old model's quality "in about half the turns, time and output tokens, cutting the cost of that workload by 40 to 50%." Viktor, which builds an AI employee that works in Slack and Teams, says it "costs nearly half as much, while getting twice as many of our hardest tasks right."
People on a subscription gain too, even though they never see a token price. Anthropic is raising the five-hour usage limits on the Pro, Max, Team and seat-based Enterprise plans. It is also giving subscribers a limit reset that they can save and use whenever they choose.
What this means for you: if your team shelved an AI project because the numbers did not work, as AI Pashut points out, this is the moment to price it again.
What it can do
Opus 5.5 is built for long, sprawling jobs, the kind that used to take a team several weeks. The headline example from Anthropic's launch: one early tester "completed a 680,000-line code migration in less than a day—work that would have taken an engineering team weeks." Another left it running overnight on a task spread across six codebases, and it "stayed on task for over 18 hours".
Much of this work is done as an agent: an AI that carries out a task step by step on its own, using the tools and files you give it, and checks its work as it goes. You set the goal; it does the steps.
You do not have to be a programmer to benefit, and the office-work results say more than the coding ones. Anthropic asked Opus 5.5, Fable 5.1 and Opus 5 to write a report on a company's quarterly results, using only a copy of the web where the key document was hard to find. An automated grader checked every figure and quote, and a single invented number failed the report. 16 of 18 reports from Opus 5.5 cleared that bar. Neither Fable 5.1 nor Opus 5 cleared it in any attempt. In plain terms, it found the facts instead of making them up.
Customers describe the same habit of looking twice. Hex, a data-analysis company, set it a question about whether parcels were late or tracking was just slow. The old model checked the delivery confirmations and called tracking healthy. Opus 5.5 "found the packages were late and tracking was broken too." At the investment firm Walleye Capital, it noticed that the minute numbering in the firm's own test instructions was off by one, and corrected for it, and the firm says "no model we've tested had caught and acted on that before."
You can use it today in the Claude apps and in Claude Code, and developers reach it as claude-opus-5-5 on the Claude Platform. It is also available on Amazon Web Services, Google Cloud and Microsoft Azure.
What this means for you: hand it the job you keep postponing because it is too big for one afternoon, such as a report built from many sources or a clean-up across a whole folder of documents, and check the result the way you would check a new colleague's work.
Why this is good news
Competition is working for the customer. Three weeks ago, top-level AI cost $10 per million tokens in and $50 out, whichever of the two leading makers you chose. Today a model that Anthropic says performs at the level of Fable 5.1 on most work costs $4 and $20. Nobody had to wait a year for that price to fall.
The good model becomes the everyday model. When the top model is expensive, teams save it for the hardest jobs and send everything else to a cheaper, weaker one. A lower price changes that habit. The software company Factory says Opus 5.5 is "the first model we'd default to at medium effort". Effort here means how long the model thinks before it answers, and medium is the everyday setting. In practice, the good model becomes the one you use without thinking about it.
Less back-and-forth, more thinking. The most human sentence in the launch comes from Quantium, a data and AI consultancy. A complex task that used to take 38 prompts over four days came in at 11 prompts over three hours, and the company describes what its people did with the time saved: "less time iterating and more time interrogating: testing assumptions, pressure-testing outputs, and landing on the best solution for our clients." The saving is not only money. It is the hours people used to spend nudging the machine, now spent judging its work.
And there is more to come. Anthropic says the two smaller models in the same family, Claude Sonnet 5.5 and Claude Haiku 5.5, "will follow in the coming weeks" with many of the same improvements. Cheaper everyday work gets cheaper again.
What this means for you: the question used to be whether you could afford the best model for a job. Now it is which job you would give it first.
What got better: clearer writing, safer on its own
It writes like a colleague. One of the most common complaints about Opus 5 was that its answers were long and hard to follow. Anthropic says Opus 5.5 "puts the most important information up front, is less likely to use jargon or idiosyncratic phrases, and follows the writing rules you give it." One early tester put it simply: "it writes the way I do." An engineer at the finance company Ramp went further: "It writes like a good colleague, and follows our writing rules."
Anthropic's launch page shows the difference with the same question put to both models. A customer's bill had dropped, and the user asked each model to explain why.
The question to both models: "Please explain the issue to me."
Opus 5 opens with
It then goes straight into the code. The money is never stated.
Opus 5.5 opens with
After the code, it ends on the business consequence: "The usage isn't moved to another month; it's never billed at all."
Anthropic makes a further point about this. Clearer writing makes the model's work "easier to follow and check—which is a safety benefit as well as a practical one." A mistake you can see is a mistake you can catch.
It is safer when it works on its own. The more an AI does without supervision, the more its judgement matters. On Anthropic's broadest behaviour test, nearly 2,000 simulated scenarios, Opus 5.5 scored better than any recent Claude model on nearly every measure of bad behaviour. Anthropic says it is much less likely to take actions that are hard to undo. In a new test of whether a model tries to get out of the boundaries it has been given, it tried about 85% less often than Opus 5, and it reported every attempt itself.
It is also harder to trick. A prompt injection is a hidden instruction planted in a web page, an email or a document, written to hijack an AI that reads it. On a test run by the security firm Gray Swan, Opus 5.5 tied Fable 5.1 for the lowest success rate of such attacks of any model tested. Anthropic is open about the limits of all this testing: it sees signs that the model "often suspects it is being evaluated", which makes a lab result an imperfect guide to real life.
Three smaller changes are worth knowing about. Most cybersecurity tasks are handed automatically to an older model, Opus 4.8, as a safeguard, and the model can no longer be run with its thinking step switched off. For readers in Europe, it carries Anthropic's watermarking measures for the EU AI Act, and it is available with zero data retention for organisations that need it.
What this means for you: try it on something you write every week, such as a status update, a summary of a long thread or an explanation for your manager, and see whether the first line is the one your reader needs.
Official sources
- anthropic.comAnthropic: "Introducing Claude Opus 5.5", the launch announcement, 22 September 2026
- anthropic.comAnthropic: Claude Opus 5.5 System Card, the full safety and testing report
- platform.claude.comAnthropic developer docs: model pricing, including Opus 5.5, Opus 5 and Fable 5.1
- darioamodei.comDario Amodei: "We Must Pace the Frontier", September 2026
- openai.comOpenAI: introducing GPT-6 Astra, 3 September 2026
- developers.openai.comOpenAI developer docs: GPT-6 Astra model page and pricing
- ai-pashut.co.ilAI Pashut (Hebrew), Daniel Nakhemia and Benny Farber: Claude Opus 5.5 and what it means for your business