What is Claude Opus 5, and how does it compare with Claude Fable 5?

Yesterday, Friday, July 24, 2026, Anthropic released Claude Opus 5. Anthropic's lineup now has a clear top tier, Claude Fable 5, and Opus 5 is meant to sit just below it as the model people can afford to use all day. Anthropic's one-line pitch is that Opus 5 gets close to Fable 5's intelligence for half the price.

It is available everywhere Claude runs: the claude.ai apps, Claude Code for programming, Claude Cowork for office tasks, and the Claude API that developers build on. On the $100-a-month Claude Max plan it becomes the default model; on the $20-a-month Pro plan it is the most capable model available.

What Anthropic released

The key facts fit in a small table:

A token is a small piece of text, roughly part of a word, and API prices are counted per million. So Opus 5 brings a large jump in ability with no change in cost compared with its predecessor, Opus 4.8.

Two developer features arrive with it, both in beta. One lets a program change which tools Claude may use in the middle of a conversation without losing its prompt cache, the stored copy of earlier instructions that makes repeated requests cheaper. The other is automatic fallback: if a safety filter flags a request, the API can quietly send it to another model instead of refusing outright.

How it performs

Anthropic publishes many scores, and they are its own measurements. On coding tests such as Frontier-Bench, it says Opus 5 beats every other model and more than doubles Opus 4.8's result at a lower cost per task. On a business-automation test from Zapier, it reports a pass rate about one and a half times the next-best model at equal cost. On OSWorld 2.0, a test of operating a computer, it says Opus 5 tops Fable 5's best score while spending a little more than a third as much.

A useful idea for reading these charts is the effort setting. You can tell Claude how hard to think, from low to max. More effort means better answers but more tokens and time. Anthropic's claim is that Opus 5 gives more performance than rivals at almost every effort level, so you can turn it down for routine jobs and still get good results.

Techstrong.ai, reporting Anthropic's comparisons, notes that Opus 5 matches or beats Fable 5 on most tests, trailing slightly in some legal and health tasks, and beats OpenAI's GPT-5.6 Sol almost everywhere except agentic coding.

How it works differently: checking its own work

The improvement Anthropic stresses most is not raw knowledge but habits. Opus 5 is described as much better at verifying what it produces and trying again until it succeeds. Anthropic gives an example from testing: asked to rebuild a machine part as a 3D model from a drawing it was not allowed to view directly, Opus 5 wrote its own small computer-vision program to extract the shape from the image's raw pixels, then built the part.

Think of the difference between a student who hands in the first draft and one who reruns every calculation before submitting. The second student is slower per draft but needs far less correction. You can encourage that habit with your instructions:

Build a one-page budget calculator in HTML for a family of four.
Before you show it to me, test it yourself with three sets of numbers,
including one where spending is higher than income, and tell me what
you checked and anything you fixed.

The last sentence asks for the checks Opus 5 is designed to run and makes them visible to you.

Safety: what is held back and why

Anthropic says its automated behavior audit rates Opus 5 its most aligned model so far, meaning the one that best follows its intended values, with the lowest rate of deceptive behavior and the strongest resistance to being tricked into misuse.

Cybersecurity gets special handling. Anthropic deliberately did not train Opus 5 on hacking tasks, yet the model improved anyway as it became generally smarter. It is now close to Anthropic's restricted Mythos 5 at finding software flaws, but far behind it at exploiting them, that is, turning a flaw into a working attack. Because of that gap, Opus 5's classifiers, automatic filters that watch for dangerous requests, are looser than Fable 5's: Anthropic expects them to step in about 85% less often. When they do, requests in Anthropic's own apps fall back to Opus 4.8 rather than failing. Security professionals in Anthropic's Cyber Verification Program get a less restricted version.

Why it matters

Price has become a central question for businesses using AI. Techstrong.ai frames the launch against company budgets strained by heavy AI use. A model that delivers most of Fable 5's ability at half the cost changes what is practical: long coding sessions, research and document work that would be too expensive with the top model become affordable. For Anthropic, it also answers OpenAI's GPT-5.6 family, launched two weeks ago with a similar efficiency message.

When should you choose Claude Opus 5 over Claude Fable 5?

For most work, start with Opus 5. Anthropic built it as the daily model, it is the default on Max, and its own tests put it close to Fable 5 at half the price. Coding, analysis, writing, research and computer-use tasks all fit, and you can lower the effort setting for simple jobs to save time and tokens.

Reach for Fable 5 when a task is at the very edge of difficulty and the small extra margin is worth paying for, or in areas such as certain legal and health questions where Anthropic's comparisons show Fable 5 slightly ahead. If you are a developer, test both on a sample of your real tasks and compare quality against cost before deciding.