Claude Fable 5 vs Claude Opus: What Anthropic’s New Coding Model Can Actually Do

On June 9, 2026, Anthropic released Claude Fable 5, and the naming alone tells you something changed. Fable is not the next Opus. It is the first model in a new “Mythos” class that sits above Opus in Anthropic’s lineup, which makes it a tier above the previous flagship rather than a sibling to it. That framing invites skepticism, because every model launch claims to be a leap. Having dug through the benchmarks and the early real-world reports, I think the leap is real, but it is real in a specific place. Fable is not a better chatbot. It is a step-change for long, hard, agentic work, above all software engineering, and the gap over Opus widens the harder the task gets. If your work does not look like that, you can save yourself some money and stop reading soon. If it does, the numbers below should get your attention.

One piece of housekeeping first, since people searching for Claude Fable will run into it: the model was briefly pulled in mid-June under a government-ordered suspension. That order was lifted around July 1, and access has been restored. Fable 5 is available now, priced at $10 per million input tokens and $50 per million output tokens. Hold that pricing in your head, because it is the hinge of the whole Fable vs Opus question.

The coding gap is not subtle

Benchmarks are easy to game and easy to overread, so I try to look for two things: a spread across independent tests, and at least one messy real-world result. Fable 5 has both.

• SWE-Bench Pro: Fable 5 scores 80.3%, the top score of any model, against 69.2% for Claude Opus 4.8. An 11-point gap on a hard software engineering benchmark is the kind of margin that used to take a full model generation.
• SWE-bench Verified: Fable reportedly lands around 95%, which is close to the point where this benchmark stops being informative because there is almost nothing left to fail.
• FrontierCode Diamond: Cognition’s benchmark for high-quality, maintainable agentic coding is the one I find most telling, because it punishes the sloppy-but-passing code that agents love to produce. Fable 5 scores 29.3%. Opus 4.8 scores 13.4%. GPT-5.5 scores 5.7%. Fable is not incrementally ahead here; it is more than double Opus and five times GPT-5.5.
• ViBench: On this end-to-end “vibe-coding” benchmark, Fable 5 is the highest-performing model tested, building working apps in less time and with fewer tokens.

That last point matters more than it looks. A model that finishes in fewer tokens partly pays for its own price premium, since output tokens are what you are billed for.

The messy real-world result comes from Stripe, which pointed Fable 5 at a 50-million-line Ruby codebase and had it complete a codebase-wide migration in a single day. Stripe’s estimate for doing the same work by hand: a team grinding for more than two months. You should treat any vendor-adjacent anecdote with some caution, but the shape of it matches the benchmarks exactly. Fable’s advantage is not in writing a clever function. It is in holding a sprawling, ugly, long-horizon task together without losing the plot.

The same pattern shows up outside coding. Fable 5 is the first model to break 90% on Anthropic’s core benchmark of complex, long-running analytical tasks, roughly a 10-point jump over Opus. It is state of the art on nearly everything Anthropic tested, from knowledge work to vision to scientific research, and the consistent finding is that the longer and more complex the task, the larger Fable’s lead. Some of the vision results are frankly a little unnerving. Fable can extract information from dense scientific figures, and it can reconstruct the source code of a web application from screenshots of it. Sit with that second one for a moment if you build web apps for a living.

The Opus safety net hiding inside Fable

Here is the detail I find genuinely fascinating, and it is the part of the Fable vs Opus story almost nobody is talking about. Fable’s headline strength is also its headline risk: cybersecurity. Mythos-class models are markedly better at discovering and exploiting software vulnerabilities, and at what Anthropic calls agentic hacking, meaning the ability to chain reconnaissance, vulnerability discovery, lateral movement, and exploit development into one autonomous campaign. That is wonderful news if you are a defender auditing your own systems and much worse news in other hands, which is presumably what got regulators interested in June.

Anthropic’s answer is a safeguard system with a twist. When you send Fable 5 a query on certain sensitive topics, the request is not simply refused. It is routed to Claude Opus 4.8 instead, and Opus answers. Anthropic says these safeguards trigger in fewer than 5% of sessions on average, so most users will rarely see them. But think about what the design implies. Opus, the model Fable just dethroned, is quietly still on duty underneath it, serving as the safe adult in the room whenever the conversation drifts somewhere the more powerful model should not go. The fully unsafeguarded sibling, Mythos 5, was never released broadly at all; it went only to a small set of vetted cyberdefenders. I cannot think of a previous flagship launch where the outgoing model was built into the new one as its conscience. It is a strange, honest piece of engineering, and it tells you Anthropic believes its own threat model.

For everyday use the practical effect is small. If you occasionally get an answer on a security-adjacent question that feels a touch more conservative, you may simply be talking to Opus without knowing it.

So should you pay for it?

Let me be blunt about the case against, because I think honesty here is more useful than cheerleading. For short, everyday questions, drafting an email, summarizing a document, explaining a concept, debugging a hundred-line script, you will mostly not notice a difference between Fable 5 and Opus. Both are past the threshold where casual tasks stop discriminating between good models. Opus 4.8, or something cheaper still, is plenty for that work, and at $10 in and $50 out per million tokens, Fable is a genuinely expensive way to ask what to cook for dinner.

The real question is whether you do the kind of work where Fable’s lead actually shows up. There is a simple test: does the task run long, does it touch many files or many steps, and does quality compound? Multi-day agentic coding sessions, large migrations and refactors, deep research across piles of papers and figures, complex analytical work that used to fall apart around step forty. If that is your week, the FrontierCode and SWE-Bench Pro numbers are not abstractions, they are hours of your life back, and the Stripe result suggests the return can be measured in months. One completed migration pays for a great deal of token spend.

My honest read: Fable 5 is the first model where “a tier above” is a description rather than marketing. Whether that tier is worth $50 per million output tokens depends entirely on whether your problems are hard enough to reach it. Most are not, and that is fine. But the more interesting thought is the one lurking in the safeguards. We now have a model so capable at offensive security that its own maker keeps the previous flagship inside it as a chaperone, and the truly unrestricted version locked away with vetted defenders. The question stopped being whether the models are good enough. It is now who we trust to hold the good ones.

Sources: Anthropic, Vellum, Morph, VentureBeat, The Decoder, Cybersecurity Dive