Claude Opus 5.5 Explained: Fable-Level AI for 40% Less - and What It Means for Your Job

On September 22, 2026, Anthropic released Claude Opus 5.5, the first model in the new Claude 5.5 family. Smaller siblings, Claude Sonnet 5.5 and Claude Haiku 5.5, are due "in the coming weeks".

The short version fits in one sentence: Anthropic says Opus 5.5 works at the level of its most capable model, Claude Fable 5.1, on most tasks - and costs about 40% less to run than the Opus 5 it replaces.

That second half is the part to watch. Every time skilled AI work gets cheaper, a few more tasks cross the line from "interesting demo" to "cheaper than a person". This article walks through what Opus 5.5 can do, what it costs, how it compares with GPT-6 Astra, and what it means for your job - without the jargon.

Claude Opus 5.5 in 60 seconds

What is a token? A token is a small chunk of text - roughly three-quarters of a word. A million tokens is about 750,000 words, or several long novels. Prices for AI models are quoted per million tokens.

The benchmarks, in plain English

Benchmarks are standard tests that let you compare AI models on the same tasks. Anthropic published nine. Here are the ones that look most like real jobs.

Terminal-Bench 4.0 is the headline. It measures engineering work done in a command line - setting up systems, fixing builds, moving data. Opus 5.5 scores 66.4%, almost nine points clear of GPT-6 Astra at 57.9%. That is the biggest lead in the whole launch. On FrontierCode, a test of hard open-ended coding problems, the race is much closer: 54.4% against 53.3%.

The test closest to an ordinary office job is GDPval.

GDPval is built from real work tasks taken from professional jobs - the reports, analyses and documents people are paid to produce. Models are rated against each other like chess players. Opus 5.5 scores 1846, more than 100 points above its own big brother, Fable 5.1, and 300 above GPT-6 Astra.

It does not win everything, and that matters too.

On OSWorld 2.1, which asks a model to finish real tasks inside desktop apps, Opus 5.5 reaches 81.8%. On AutomationBench, which runs multi-step business workflows, GPT-6 Astra is still just ahead at 41.4% to 40.0%. And Anthropic itself warns that at this level, benchmark margins are a less reliable guide than they used to be.

The real-world tests matter more than the scores

Scores tell you who wins a test. The examples Anthropic and its early testers shared tell you what changes on a Monday morning.

• A 680,000-line code migration in under a day. One early tester moved a codebase of that size in a single working day.
• A 200,000-line codebase audited and fixed in under three hours. Opus 5 needed more than 20 hours for the same job, and 2.5 times as many tokens.
• Research reports that pass review. In an internal test, 16 of 18 research reports written by Opus 5.5 cleared Anthropic's quality bar. Neither Fable 5.1 nor Opus 5 cleared it in any attempt.
• A merger financial model in 63 minutes instead of 93, at half the cost.
• Clearer writing. Anthropic says Opus 5.5 puts the key point first. One early tester's verdict: "it writes the way I do."

Look at the first panel again. A job that took the previous model more than 20 hours now takes an afternoon. When the time drops that far, a manager stops asking "can AI help with this?" and starts asking "do we still need to staff this?"

"I'm struggling to find anything negative to say." - Sean Heintz, Clio

The price drop is the real story

The sticker price fell by a fifth. The real saving is bigger, because Opus 5.5 also finishes tasks in fewer steps: Anthropic puts the total at about 40% cheaper than Opus 5 on typical work, and its HAProxy translation test came in 51% cheaper than Fable 5.1.

Here is why that matters for careers. Companies do not automate a task because a model *can* do it. They automate it when the model does it cheaper than a person, reliably. Every price cut moves more tasks over that line - without the model getting any smarter.

Claude Opus 5.5 vs GPT-6 Astra

GPT-6 Astra, released by OpenAI on September 3, was the model to beat. Here is how the two compare on the tests Anthropic published.

The honest summary: Opus 5.5 is the stronger and cheaper choice for most engineering and office work; Astra keeps an edge on science and some automation tasks. For your career the more important point is that two labs are now neck and neck at the top - so companies can switch between them, and prices keep falling.

Is Claude Opus 5.5 safe?

Anthropic published more safety detail than most launches carry:

• It scored the best of any model Anthropic has tested on its automated behaviour audit.
• In a new test, it tried to get around its limits about 85% less often than Opus 5.
• It often suspects when it is being tested, which Anthropic flags because it makes testing harder.
• Most cybersecurity tasks are routed to an older model, Opus 4.8, as a precaution.

What Claude Opus 5.5 means for your job

Read the exposure off the evidence, not the hype.

Notice the pattern. The exposed work all has a clear finish line: the code compiles, the report passes review, the model balances. The protected work is the part where someone has to decide what "done" means, and answer for it.

What to do this month

1. List your week. Write down the five tasks that take most of your time. Mark each one that has a clear finish line. If that is more than half your week, the table above is about you.
2. Use the tool on your own work. Try Opus 5.5, or a model like it, on one real task. You learn its limits in an hour, and "knows how to direct AI" is fast becoming a hiring filter.
3. Move up a level. The model executes. You decide what is worth doing, check the result, and own it with the people who care. Those three jobs are not on any benchmark.
4. Get your number. Know which of your tasks are exposed and which strengths protect you - then build a plan around them.

Claude Opus 5.5 is not the end of anyone's career. It is another notch on a dial that keeps turning: the same work, done faster, for less. The people who do well are the ones who notice the dial moving - and move first.