Claude Opus 5 vs Fable 5: Which One Should You Actually Use?

Use Claude Opus 5. It costs $5 per million input tokens and $25 per million output, exactly half of Fable 5's $10/$50, and it beats Fable 5 on most benchmarks the two models share. Fable 5 keeps a narrow lead in one place: long unattended agent runs where a failed attempt costs more than the token bill.

That's the short version. Now the part that actually matters, because "just use the cheaper one" is useless advice when you're the person signing off on either a blown budget or a botched migration.

What's the actual difference between Opus 5 and Fable 5?

Opus 5 shipped July 24, 2026. Fable 5 shipped June 9. Both give you a 1M-token context window and up to 128K output tokens per request. Same tokenizer, same Messages API, same tool surface. The real gap is price, autonomy on very long runs, and a handful of API restrictions that only apply to Fable.

Neither charges a long-context premium, which is worth saying out loud because plenty of frontier models still tack one on past 200K.

Which one wins on benchmarks?

Opus 5 wins most of them. Anthropic's launch numbers put it ahead on Frontier-Bench v0.1, ARC-AGI-3, GDPval-AA v2, OSWorld 2.0, and DeepSearchQA. Fable 5 holds a sliver of a lead on SWE-bench Pro, and that's roughly the whole list.

Two caveats, because benchmark tables lie by omission. First, these ran on Anthropic's own scaffolding, so a one-point SWE-bench Pro gap sits inside the noise floor. Call it a tie. Second, Anthropic listed the evaluations it ran for Fable 5 without publishing raw scores on the model page, so several Fable numbers above come from third-party trackers rather than the vendor. Weight them accordingly.

The number that actually moves a decision is OSWorld 2.0. Opus 5 beat Fable 5's best computer-use result at about a third of the cost per task. That isn't a marginal win, it's a different budget line.

Where does Fable 5 still beat Opus 5?

Long-horizon autonomy. Anthropic's framing is that Fable 5 works autonomously for longer than any previous Claude model, holding focus across millions of tokens. Early testers described it compressing months of engineering into days on complex, multi-day coding tasks.

That's a real capability and it doesn't show up cleanly in any single-turn benchmark. If your agent runs overnight, delegates to sub-agents, and nobody reads the output until morning, Fable 5's extra reliability can be cheaper than Opus 5 plus two failed reruns plus an hour of your time untangling what went sideways.

The flip side: most people don't run overnight agents. They run a coding assistant, a review pass, a research task, or a chat. For all of those, the extra spend buys nothing you can measure.

Fable 5 has one platform advantage too. It supports Priority Tier; Opus 5 doesn't. Opus 5, Sonnet 5, and Mythos 5 are all excluded, and a Priority Tier request naming one of them fails validation. If you've built capacity guarantees around it, that's a genuine blocker and not a benchmark question at all.

What the API won't let you do on Fable 5

This is the section nobody writes and everyone hits in week two. Fable 5 carries real API restrictions Opus 5 doesn't:

• You cannot turn thinking off. Sending thinking: {"type": "disabled"} returns a 400. Omit the parameter entirely and control depth with output_config.effort. Opus 5 also thinks by default, but you can disable it at effort high or below.
• You cannot see the reasoning. The raw chain of thought is never returned. Setting display: "summarized" gets you a readable summary and that's the ceiling.
• No assistant prefill. Last-turn prefills return a 400 on both models, so use output_config.format for structured output instead.
• 30-day data retention is mandatory. Fable 5 isn't available under zero data retention. If your org is set below that, every request returns a 400 with a perfectly valid payload. I've watched someone burn an afternoon on that before anyone thought to check the retention config.
• Safety classifiers can decline a request. You get HTTP 200 with stop_reason: "refusal" and empty content. Code that reads content[0] without checking stop_reason first will break. Opus 5 ships extra cybersecurity safeguards too, so handle it on both, but Fable's classifiers are tuned harder on bio and cyber content.

Opus 5 picks up two things Fable doesn't have. Fast mode (speed: "fast", roughly 2.5x output speed at double the price, Claude API only) and a prompt cache minimum of 512 tokens instead of 1,024. That second one is quietly useful: prompts that were too short to cache on Opus 4.8 now cache on Opus 5 with zero code changes.

One more gotcha worth knowing before you shift volume. Opus 5 sits in its own rate-limit bucket and does not draw from the shared Opus 4.x pool. Moving traffic over doesn't inherit your old headroom.

Which one do you get on your Claude plan?

Depends what you pay, and the answer changed on July 20, 2026.

The "50% of limits" line trips people up constantly. It doesn't mean extra usage. Fable draws from the same weekly pool and burns through it faster, which is exactly what you'd expect from a model costing twice as much per token.

So which should you actually pick?

Pick Opus 5 if you're doing any of this:

• Interactive coding in Claude Code, Cursor, or whatever editor you live in
• Code review, debugging, bug hunting
• Computer use and browser automation, where the OSWorld gap is genuinely large
• Research, extraction, summarization, document work
• Anything cost-sensitive at volume
• Anything running under zero data retention

Pick Fable 5 if:

• You run multi-hour or multi-day autonomous agents with nobody watching
• A wrong answer costs more than the token bill (compliance work, financial modeling, migrations you get one shot at)
• You need Priority Tier capacity guarantees
• You've benchmarked both on your own eval set and Fable actually won

That last bullet is the honest one. Vendor benchmarks are directionally useful and nothing more. Run both against twenty real tasks from your own workload before you commit either way.

What I'd do

Default everything to Opus 5 at high effort, then sweep downward. Opus 5 is unusually strong at low and medium, and that's the biggest cost lever most teams never pull. Effort settings carried over from Opus 4.8 are almost certainly wrong here, so re-tune rather than assuming they transferred.

Keep Fable 5 as the escalation path. Route overnight jobs to it, keep everything interactive on Opus 5, revisit in a quarter. Anthropic has shipped four models in under two months. Whatever you pick today is a temporary decision, and treating it as permanent is the real mistake.

If you're migrating from Opus 4.8 or earlier, it's mostly a model-ID swap. Two things will bite you. Thinking is now on by default, and max_tokens caps thinking plus response text together, so a tightly sized limit will truncate mid-answer. And pairing disabled thinking with xhigh or max effort returns a 400, validated per request, so audit every call site rather than just the first one.

For launch-day detail on Opus 5, we covered pricing, fast mode, and what changed and the pre-launch leaks and specs. If you want the geopolitics instead of the benchmarks, Fable ended up at the center of a White House accusation against Moonshot AI. And for a non-Anthropic take on frontend work, Kimi K3 versus Claude is worth twenty minutes.

Official announcements