Opus 5 vs Fable 5: which Claude should you actually reach for?
Opus 5 ($5/$25) matches much of Fable 5 ($10/$50) at half the price — within 0.5% on CursorBench 3.2, ahead on OSWorld 2.0 — and is now the default in Claude Code. Fable 5 is Mythos-class and still leads the hardest frontier tail. Route by task: Opus 5 for most work, Fable 5 for the extreme tail.
As of July 24, 2026: Claude Opus 5 (
claude-opus-5, $5/$25) just shipped as the new default in Claude Code. Claude Fable 5 (claude-fable-5, $10/$50) is Anthropic’s Mythos-class model, a tier above Opus. The interesting question isn’t “which is better” — it’s “which earns its price on your workload.” Here’s my honest, sourced head-to-head.
The one-sentence answer
Opus 5 is the new value default; Fable 5 is the frontier ceiling you pay a premium to reach. For the vast majority of agentic coding and computer-use work, Opus 5 now gets you within a rounding error of Fable 5 at half the cost — which is exactly why Anthropic made it the default in Claude Code. You reach for Fable 5 when you’re on the genuinely hardest tail of tasks, where its Mythos-class ceiling still pulls ahead and you’ve decided that extra headroom is worth 2x.
Everything below is me showing my work on that claim.
Are they even the same tier?
No — and that’s the thing to internalize before you compare scores. Fable 5 is Mythos-class, which Anthropic positions “above our Opus class in capability.” Opus 5 is the top of the Opus class. So structurally, Fable/Mythos still sits a rung higher.
What changed on July 24 is how much of that rung Opus 5 closes. Anthropic’s own line is that Opus 5 “matches many of Fable 5’s capabilities” — at half the price. So the gap didn’t vanish; it got narrow enough that, for most work, paying double for the higher tier stopped making sense. The tier hierarchy is intact. The economics of using it flipped.
The benchmarks: the fair head-to-head
I’m only going to compare numbers Anthropic actually put side by side, because that’s the honest way to do this.
- CursorBench 3.2: Opus 5 “performs within 0.5% of Fable 5’s peak score, but at half the cost.”
- OSWorld 2.0 (computer-use / desktop agents): Opus 5 is “surpassing Fable 5’s best result at just over a third of the cost.”
That’s the whole story in two lines: on real agentic-coding and computer-use benchmarks, Opus 5 essentially ties or beats Fable 5 while costing a half to a third as much.
Opus 5 vs Fable 5: head-to-head (Anthropic's own numbers)
Where Fable 5 keeps its edge is the absolute frontier. It launched #1 on the Artificial Analysis Intelligence Index and posts 80.3% on SWE-Bench Pro — the credential of a Mythos-class model. Anthropic didn’t publish an Opus 5 SWE-Bench Pro figure alongside it, so I won’t invent a comparison; the honest read is that on the hardest codebase-wide tail, the higher tier still leads. On its own turf, Opus 5 more than doubled Opus 4.8 on Frontier-Bench v0.1 (43.3% vs 18.7%) and reached 68.8% on DeepSWE v1.1 — strong, and enough to make it the sensible default, without claiming the crown.
What each one costs
This is where the decision usually gets made.
Output price per 1M tokens (list)
Opus 5 is $5 / $25 per million input / output tokens. Fable 5 is $10 / $50 — exactly double. Fable 5 does bring a 1,000,000-token context window, which is a real reason to pick it when a single task genuinely needs to hold a huge codebase or document set in one window. But for most work, you’re paying 2x for a tier you close most of the gap to at Opus 5 prices. For how token prices become a monthly bill, see how much it costs to add AI to your app.
The caveats that come with Fable 5
Price isn’t the only tax on the higher tier. Two things ship with Fable 5 that Opus 5 doesn’t carry:
Mandatory data retention. Fable 5 came bundled with a 30-day retention policy that applied even to customers with zero-retention agreements. Opus 5 has no data-retention requirement for general access. If you operate under zero-retention commitments, that’s not a footnote — it can be the whole decision.
The classifier reroute. On flagged topics (cybersecurity, bio/chem, distillation), Fable 5 hands the turn to Opus 4.8 and tells you. It’s rare (Anthropic says 95%+ of sessions never see it), but it means on those topics you can pay Fable prices for Opus-tier output. Opus 5’s cyber classifiers are “proportionally less restrictive than those on Fable 5,” and its announcement doesn’t describe that automatic reroute — so security-adjacent debugging should feel smoother on Opus 5. (Deliberately, Opus 5 stays behind Mythos 5 on offensive-security tasks; that’s by design, not a gap to fix.)
Where each one genuinely wins
Where each model genuinely wins
Claude Opus 5 ($5/$25)
- The new everyday default (incl. Claude Code)
- CursorBench within 0.5% of Fable at half cost
- OSWorld above Fable at ~1/3 cost
- No mandatory data retention
- /effort dial to tune cost vs quality
Claude Fable 5 ($10/$50)
- Mythos-class ceiling, above Opus tier
- Hardest frontier tail (SWE-Bench Pro 80.3%)
- 1,000,000-token context window
- Longest-horizon autonomy (Stripe: 50M-line migration in a day)
- Worth 2x only when the tail earns it
FAQ: quick answers builders ask
Which is the default in Claude Code? Opus 5, as of July 24, 2026.
Is Opus 5 as good as Fable 5? On many tasks, within a rounding error — CursorBench within 0.5%, OSWorld ahead — at half to a third of the cost. On the hardest frontier tail, Fable 5 (Mythos-class) still leads.
When is Fable 5 worth the 2x? Genuinely hard, long-horizon, codebase-wide work; tasks that need the 1M context window in one shot; or when you’ve measured a real quality gap on your own evals. My deeper take: is Claude Fable 5 worth it.
What about retention and reroute? Fable 5 carries mandatory 30-day retention and a flagged-topic reroute to Opus 4.8. Opus 5 has neither for general access.
How do I keep Opus 5 cheap? Use /effort — High or below for routine work, Extra/Max only for the hard problems. More in Claude Opus 5, explained.
The verdict: route by task, not by hype
Route by task shape
Here’s how I’d actually operate it. Make Opus 5 your default — Anthropic already did, in Claude Code — and let it handle the bulk of your agentic work at $5/$25. Keep Fable 5 in your back pocket for the hard tail: the frontier-difficulty tasks, the ones that need a million-token window in one shot, the cases where you’ve measured a real gap on your own evals. That’s not fence-sitting; it’s just refusing to pay 2x by default for headroom most tasks never touch.
If you want a second set of eyes on the routing — which of your workloads truly need Fable 5, and where Opus 5 quietly saves you half — that’s the kind of work I do. And if you’re standing up agents on either model, start with building an autonomous agent on Claude.