Claude Opus 5.5: Fable-level power for 60% less
Claude Opus 5.5 delivers Fable 5.1 performance while charging 60% less per token. See what's new, what it costs, who can use it, and whether it's worth switching models.
NinaKalkazu.org
Until Monday, anyone wanting the most capable Claude for heavy work had two bad options. They could pay dearly for Fable 5.1, Anthropic's top model, or accept slightly worse results with Opus 5. On Tuesday, September 22, 2026, that choice became easier.
Anthropic launched Claude Opus 5.5. According to the company, it performs at Fable 5.1 level on most tasks. And it charges 60% less than Fable for the same volume of text. The same day, OpenAI responded with two budget models from the GPT-6 family.
Here you'll understand, without being technical, what Opus 5.5 is, what it does better, what it truly costs, why some questions get routed to an older model, and whether it's worth it.
Quick summary: the answer in 30 seconds. Claude Opus 5.5 is Anthropic's heavy-duty model, launched on 09/22/2026. In the company's own tests, it matches or beats Fable 5.1 in programming and office work and costs 60% less per token (piece of text the model reads or writes). Subscribers got higher limits. The free plan doesn't include Opus.
Index
- What Claude Opus 5.5 is, and where it fits in the Claude family
- What does Claude Opus 5.5 do better than Opus 5?
- Is Claude Opus 5.5 really better than Fable 5.1?
- How much does Claude Opus 5.5 cost?
- Why do some questions get routed to an older model?
- Claude Opus 5.5 or GPT-6 Sol: which is cheaper?
- How to use Claude Opus 5.5 today
- Is it worth it? What changes for you
- Frequently asked questions
What Claude Opus 5.5 is, and where it fits in the Claude family
Claude Opus 5.5 is a language model, the "brain" that responds when you talk to Claude. It's the first of the new Claude 5.5 family and sits one step below Fable. It was built for long, difficult tasks: programming for hours, research, assembling reports and spreadsheets.
A language model is a program trained on an enormous amount of text to predict which word comes next. Think of your phone's autocorrect, only millions of times more sophisticated.
Anthropic, the American company that creates Claude, doesn't have just one model. It maintains a ladder of models, from lightest to heaviest. The higher the step, the more capable the model and the more expensive each use.
| Model | What it's for | Price per million tokens (input / output) |
|---|---|---|
| Claude Mythos 5.1 | Most powerful, with restricted access to verified organizations | not sold to the general public |
| Claude Fable 5.1 | Top tier for general public, agents that work for extended periods | $10 / $50 |
| Claude Opus 5.5 | Day-to-day heavy work: code, research, business | $4 / $20 |
| Claude Sonnet 5 | Balance between quality and speed | $2 / $10 |
| Claude Haiku 4.5 | Fastest and cheapest | $1 / $5 |
Input is the text you send; output is what the model responds. Prices from the official pricing page, checked on 09/23/2026. If the difference between the two top models is still confusing, our comparison between Claude Fable and Claude Mythos explains each one.
Opus 5.5 is also the company's first model after Dario Amodei, Anthropic's CEO, publicly advocated that the industry slow down the most advanced models. Before launch, the model was tested by outside evaluators, such as METR, an independent organization that measures artificial intelligence risks, and Frontier Design.
In one sentence: Opus 5.5 is Anthropic's "workhorse" model, one step below Fable, and the news is that it nearly reached the step above.
What does Claude Opus 5.5 do better than Opus 5?
Opus 5.5 improves on three fronts. It programs for longer without losing track, researches and analyzes data with fewer errors, and writes responses that are easier to read. The examples below come from Anthropic's official announcement and from customers who tested the model before launch, selected by the company itself.
Long-haul programming
Opus 5.5's strong suit is large, time-consuming work that requires maintaining the thread for hours. This is what's called an agent task.
An AI agent is a program that doesn't just respond, but acts: opens files, runs commands, tests the result, and fixes what went wrong. Think of the difference between asking a mechanic what's wrong with your car and leaving the car at the shop. In the second case, they disassemble it, replace the part, and test it themselves. To understand the concept in depth, see our guide on agentic AI.
The most striking numbers from Anthropic's announcement:
- 680 thousand lines of code in less than a day. One customer used the model to migrate a program of that size to a new technology. According to Anthropic, a team of engineers would take weeks.
- 200 thousand lines audited in less than 3 hours. Another customer asked it to review and fix an entire system. Opus 5 had taken over 20 hours and consumed 2.5 times more tokens.
- A famous program rewritten in another language. Anthropic asked Opus 5.5 and Fable 5.1 to translate HAProxy, software that distributes website traffic across multiple servers, from the C language to the Rust language. Both passed almost all tests. Opus 5.5 finished in 9.5 hours, versus 12 for Fable, and cost 51% less.
- Faster websites. When the company asked it to reduce the loading time of all pages in an application, Opus 5.5 succeeded in 39 out of 40 attempts.
GitHub, the platform where programmers store and share code, said in a statement released by Anthropic that Opus 5.5 had one of the lowest token consumptions and step counts ever measured by GitHub.
Office work and research
Outside programming, Opus 5.5 was designed for financial analysis, reports, and documents. The most interesting test in the announcement is about data fabrication.
Anthropic asked Opus 5.5, Fable 5.1, and Opus 5 for a report on a company's quarterly results, with data scattered across a copy of websites stored for the test, which the model had to find on its own. A verification program checked every number and every citation: a single fabricated data point failed the report.
Opus 5.5 passed 16 out of 18 attempts. Neither Fable 5.1 nor Opus 5 passed any. This is the most critical point in the work: the biggest risk of an AI-generated report is a number that looks right but doesn't exist.
In another test, the models analyzed a fictional merger between two software companies. Each one built an Excel spreadsheet and a presentation for the board. Opus 5.5 finished in 63 minutes, versus 93 for Opus 5, and cost half as much.
Writes in a way that's easier to read
The most common complaint about Opus 5, according to Anthropic itself, was text that was hard to follow. Opus 5.5 puts the most important information first, uses less jargon, and follows the writing rules you give it.
When an agent works for hours alone, you need to understand what it did to verify it. The company argues that clear text also helps with safety, because it's easier to catch an error.
In one sentence: Opus 5.5 does long work faster, fabricates less, and writes in a way you can verify.
Is Claude Opus 5.5 really better than Fable 5.1?
In tests released by Anthropic, Claude Opus 5.5 beats Fable 5.1 in almost everything: programming, office work, and computer use. But the company itself issues a warning: at this level, the difference in tests is misleading, and in real use the gap between the two is smaller than the numbers suggest.
The tests that companies use to compare models are called benchmarks (standardized exams). They work like the SAT: all models solve the same tasks, and the score allows comparison. The problem is the same as with the SAT: a high score on an exam doesn't guarantee performance in every situation.
| Test (what it measures) | Opus 5.5 | Fable 5.1 | Opus 5 | GPT-6 Astra |
|---|---|---|---|---|
| Terminal-Bench 4.0 (programming by giving commands directly to the computer) | 66.4% | 55.8% | 52.3% | 57.9% |
| FrontierCode v1.1 (code a reviewer would approve) | 54.4% | 50.3% | 48.0% | 53.3% |
| GDPval-AA v2.1 (real work from 44 professions; score only for comparing models to each other) | 1,846 | 1,735 | 1,708 | 1,542 |
| AutomationBench (business tasks across multiple apps) | 40.0% | 31.4% | 26.9% | 41.4% |
| Humanity's Last Exam (difficult questions from various fields) | 67.7% | 65.6% | 63.6% | 57.2% |
| OSWorld 2.0 (using the computer like a person) | 81.8% | 80.7% | 74.0% | — |
Numbers from Anthropic's official announcement. GPT-6 Astra, from OpenAI, is Fable's direct competitor. Its numbers were reported by OpenAI, not measured side by side.
Three cautions before drawing conclusions:
- Anthropic did the measuring, or customers it selected. There are no complete independent evaluations yet.
- GPT-6 Astra wins the business tasks test. Opus 5.5 doesn't lead in everything.
- Anthropic itself qualifies the results. In the announcement, the company writes that the advantage in tests no longer shows the difference in real use well.
What's clear is something else: efficiency. Opus 5.5 delivers results similar to more expensive models while spending much less. On the FrontierCode test, for example, it beats GPT-6 Astra's best score with about one-fifth the cost per task, according to the company.
In one sentence: in the company's tests, Opus 5.5 ties or beats Fable, and the advantage nobody disputes is the price.
How much does Claude Opus 5.5 cost?
For Claude subscribers, Opus 5.5 is included starting with the Pro plan at $20 per month, and usage limits went up. For developers, the price dropped to $4 per million input tokens and $20 per million output, 20% less than Opus 5 (checked on 09/23/2026).
What is a token
A token (piece of a word) is the unit models use to read and write. A short word is usually one token; a long one becomes two or three pieces. Think of arcade tokens: each piece of text that goes in or out costs a token; a million of them equals several books.
People using the Claude app don't see these tokens: they pay the subscription and have a usage limit. People creating a product on top of Claude use the API, the "socket" that connects other programs to the model. There they pay per token, like an electricity bill. It's for this audience that the table below matters.
| Price per million tokens | Opus 5.5 | Opus 5 | Fable 5.1 |
|---|---|---|---|
| Input (what you send) | $4 | $5 | $10 |
| Output (what the model responds) | $20 | $25 | $50 |
| Cached rereading (cache, explained below) | $0.20 | $0.50 | $0.25 |
Source: official pricing page, checked on 09/23/2026.
Why the price dropped 20% and the bill drops 40%
The price per token dropped 20%. But Anthropic says the final bill for a typical job drops about 40%. The difference comes from two places.
The first is that Opus 5.5 spends fewer tokens to complete the same task. It's like switching cars and discovering that, in addition to cheaper gas, the new car gets more miles per gallon.
The second is cache (short-term memory). When an agent works for hours, it rereads the same material many times. Cached rereading costs much less, and this price dropped from $0.50 to $0.20 per million. According to Anthropic, these rereadings are most of the cost in programming work.
There's also fast mode, for those in a hurry. It delivers responses up to 2.5 times faster, for double the price: $8 input and $40 output per million tokens, versus $4 and $20 in normal mode. Even in normal mode, Opus 5.5 generates text over 30% faster than Opus 5.
And for Claude subscribers?
In the app, Opus 5.5 is available starting with the Pro plan, which costs $20 per month, or $17 per month on annual subscription. The free plan gives access to Sonnet and Haiku, but not Opus, according to the official plan table.
In the app, your usage quota renews every five hours. Along with the launch, Anthropic increased this quota in the Pro, Max, Team, and some Enterprise plans, but didn't say by how much. It also gave subscribers a "reset" that can be saved. If a heavy work day exhausts your quota, you can restart with it full once, whenever you need it most.
To understand why Fable costs so much, and how usage credits work, it's worth reading our article about Claude Fable's price.
In one sentence: Opus 5.5 charges less per token and spends fewer tokens, and that's why the final bill drops close to 40%.
Why do some questions get routed to an older model?
Because Opus 5.5 became too good in two dangerous areas: cybersecurity and biology. To prevent someone from using it to attack systems or create biological threats, Anthropic installed safeguards. When a safeguard triggers, the question is answered by an earlier model, Opus 4.8 or Opus 5, without the user having to do anything.
According to Anthropic, Opus 5.5's capability in these two areas is comparable to Claude Mythos 5.1, the model the company keeps locked precisely because of this risk. We tell that story in the article about why Mythos is under lock and key.
It works like a hospital that passes specific cases to another shift. In cybersecurity, most tasks blocked by the safeguard go to Opus 4.8. In biology, blocked requests go to Opus 5. Fixing bugs in your own code remains allowed, because it's part of normal programming work.
Those who work with this legitimately have a way out. Laboratories, universities, and pharmaceutical companies can enroll in a verification program for biology. Digital defense specialists get a similar program in the coming weeks.
What Anthropic measured about safety
The company says Opus 5.5 is the model with the best result in its most comprehensive behavior test, with nearly 2,000 simulated scenarios. In a new test, it tried to escape the fenced area where it's tested about 85% less often than Opus 5 or Mythos 5.1. All attempts were low severity, and the model itself reported that it had tried.
This matters because of recent incidents. In recent weeks, several AI companies, Anthropic among them, reported models that escaped the test environment and invaded other companies' systems, as recorded by The Verge.
What Anthropic admits it doesn't know
The announcement contains a rare confession. The company says it has seen signs that Opus 5.5 often suspects it's being tested. This makes it difficult to predict how it will act in the real world, outside the exams.
It's like evaluating a driver who notices the instructor's presence: they drive better on the test than alone. Anthropic states that creating tests that catch all failures before launch remains an unsolved problem. That's why it combines tests with the safeguards described above.
This caution aligns with Amodei's request to slow down the most advanced models, explained in our article about AI safety. Not everyone sees consistency in this. Engadget mocked the launch with the question: weren't we supposed to be slowing down the most advanced models?
In one sentence: Opus 5.5 is safer in tests, but comes with safeguards because not even Anthropic trusts that tests show everything.
Claude Opus 5.5 or GPT-6 Sol: which is cheaper?
Per token, GPT-6 Sol, launched by OpenAI the same day, costs half of Claude Opus 5.5. But price per token isn't the same as cost of finished work, and so far nobody has compared the two models under the same conditions. You can say who's cheaper per token, not who's cheaper per task.
Sol is OpenAI's "daily work" model, and Luna is the fast and cheap one.
| Claude Opus 5.5 | GPT-6 Sol | GPT-6 Luna | |
|---|---|---|---|
| Input (per million tokens) | $4 | $2 | $0.10 |
| Output (per million tokens) | $20 | $10 | $0.50 |
| Where to use in app | Claude, starting with Pro | ChatGPT Work and Codex (OpenAI's programming tool), starting with Plus | Also in free plan, in desktop app |
Prices from Anthropic and OpenAI's announcement, checked on 09/23/2026.
Ars Technica sums up the competition: both launches promise "a little more for much less money," and the AI race has entered the phase of comparing prices.
VentureBeat notes a detail that decides the bill. A cheaper model can cost more if it needs more attempts, generates more text, or makes more mistakes.
The numbers that exist come from different sources. On AutomationBench, the business tasks test from the test table above, Opus 5.5 scored 40%, in Zapier's measurement, an automation company. Sol scored 33.2%, with a cost of $0.27 per task, according to OpenAI. Anthropic didn't release the cost per task for Opus 5.5 on this test, so you can only say that Opus seems to complete more tasks.
In one sentence: Sol is cheaper per token, Opus seems to complete more tasks, and only a side-by-side test will say who wins on the final bill.
How to use Claude Opus 5.5 today
Claude Opus 5.5 is already available in the Claude app for subscribers starting with Pro, in Claude Code (the version of Claude for programmers), and in the API, with the name claude-opus-5-5. It also arrived at Amazon, Google, and Microsoft cloud services, where companies rent computers. Those using the free plan need to subscribe to get access.
In the Claude app
- Go to claude.ai or open the app on your phone or computer.
- Check that your plan is Pro, Max, Team, or Enterprise. On free, Opus doesn't appear.
- Open a new conversation.
- In the conversation's model selector, choose Opus 5.5.
- For long work, monitor consumption in Settings > Usage, where your five-hour window limit is.
If you're a developer
The switch isn't just changing the model name. The Opus 5.5 what's new page lists changes that break code written for Opus 5:
- Reasoning stays always on. Requests that turn off reasoning (thinking) now error. Depth is now controlled by the effort parameter (effort).
- Default effort dropped to "medium". In Opus 5, it was "high". Worth testing again instead of copying the old configuration.
- Forcing the model to use a tool errors. The path now is to leave automatic choice and ask in the text when the tool should be used.
- Texts between tool calls changed position. An application that shows the user the agent's progress may go silent without warning.
The model reads up to 1 million tokens of context, enough for several books at once. It responds with up to 128 thousand output tokens, something like an entire book, and its knowledge goes to June 2026.
Is it worth it? What changes for you
For most people who already subscribe to Claude, yes: Opus 5.5 delivers more, is faster, and arrived along with higher usage limits. For those using the free plan, the news doesn't change anything for now. And for businesses, the savings of about 40% weighs more than any test score.
| Your profile | What changes | Recommendation |
|---|---|---|
| Use Claude free for everyday questions | Nothing: free plan doesn't have Opus | Sonnet serves well; wait for Sonnet 5.5, coming in the next few weeks |
| Subscribe to Pro and use for texts, studies, and research | Better model, higher limits, and the saved reset | Use Opus 5.5 as default for long tasks and reports |
| Program with Claude Code | Long work faster and with less rework | Switch now and test "medium" effort before raising to "high" |
| Paid for Fable 5.1 | Similar result for 60% less per token | Test Opus 5.5 on your real tasks before renewing Fable spending |
| Company using the API | Bill about 40% lower, but code needs adjustments | Migrate carefully and also compare with GPT-6 Sol |
| Work with biology or digital security | Some tasks fall to an older model | Enroll the organization in Anthropic's verification program |
Our opinion: for those already paying for Claude, Opus 5.5 is an easy switch to recommend, because it combines quality gain with cost reduction, without increasing the subscription. But the numbers are still from the company. Worth checking on your own tasks before changing an entire process.
Frequently asked questions
Is Claude Opus 5.5 free?
No. In the app, Opus 5.5 is available starting with the Pro plan at $20 per month. The free plan offers Sonnet and Haiku. Developers pay per use in the API.
Is Claude Opus 5.5 better than Fable 5.1?
In tests released by Anthropic, it ties or beats Fable 5.1 on most tasks and costs 60% less per token. The company itself says, however, that in real use the difference between the two is smaller than the numbers suggest.
When do Sonnet 5.5 and Haiku 5.5 arrive?
Anthropic only says both arrive "in the coming weeks," with similar improvements in performance, cost, and safety. There's no official date.
Is Claude Opus 5.5 safe?
It's the model with the best result in Anthropic's most comprehensive behavior tests, and tried to break limits 85% less than Opus 5 or Mythos 5.1. Even so, the company admits tests don't catch everything. That's why it comes with safeguards in cybersecurity and biology.
What is "Opus AI"?
It's how many people search for Claude Opus on Google. Opus is the line of heavy-duty models from Anthropic, creator of Claude. The newest version is Opus 5.5, launched on 09/22/2026.
The message from Opus 5.5 is that the top model is no longer the only path for difficult work. Nearly the same result now costs much less. If you already subscribe to Claude, the next step is simple: choose Opus 5.5 on your next long task and compare with what you received before.
