Claude Opus 4.8 vs. Claude Fable 5: A Practical Comparison (July 2026)

For eleven days at the end of May, Claude Opus 4.8 was arguably the most capable model you could buy. Then, on June 9, 2026, Anthropic released Claude Fable 5, its first publicly available model from the new Mythos tier, sitting a rung above Opus, and the ground shifted. If you build with these models, or pay for them, the practical question isn't "which is better." Fable 5 is better on the benchmarks. The real question is narrower and more useful: better enough to justify twice the price, for the work you actually do?

I've spent enough time routing real workloads between capable and cheaper models to be skeptical of launch-day leaderboards. They measure the ceiling; most work happens well below it. So this comparison is organized around that gap: where Fable 5's advantage is decisive, where it evaporates, and how to decide per task rather than per hype cycle.

A note on sourcing: the benchmark figures below come from Anthropic's launch materials and independent trackers. I've flagged what's vendor-reported versus third-party, because on a fast-moving model release that distinction matters.

The one-paragraph answer

Fable 5 wins on hard, long, autonomous work: big refactors, multi-step agents, complex builds, and the harder the task, the wider the gap. Opus 4.8 wins on everyday work, cost, and predictability: it's half the price, has no safety-routing fallbacks, and on short tasks performs within a hair of Fable 5. If your bottleneck is capability, pay for Fable 5. If your bottleneck is your invoice, or you're doing routine work, Opus 4.8 is the smarter default. Most teams should use both.

What the benchmarks say

The headline numbers, cross-checked against Anthropic's June 9 launch table and third-party trackers:

Read the gaps, not just the winners. On verified software tasks the difference is meaningful but modest (+6.4). On the hardest coding split it more than doubles (2.2x). On knowledge work it's a small edge. On computer use, nearly a tie. That pattern is the whole story: Fable 5's lead scales with difficulty, length, and autonomy. The longer and more self-directed the task, the more its per-step advantage compounds across steps.

The number that actually changes the decision

Sticker price says Fable 5 is the premium tier you save for special occasions: $10/$50 per million tokens versus $5/$25 for Opus 4.8, exactly double. But tokens are the wrong unit. Completed tasks are.

Here's the finding that broke my own mental model: on SWE-bench Pro, Fable 5 at low effort scores about 75, while Opus 4.8 at extra-high effort scores about 69, and Fable 5 at low effort costs less per completed task than Opus at maximum effort. The community shorthand, "Fable 5 on low is about Opus 4.8 on high," turns out to be roughly true on hard agentic coding.

The implication is counterintuitive and worth sitting with: on genuinely hard problems, the more expensive model can be cheaper per success, because it solves in fewer turns with fewer failed attempts. Cost-per-token is what the pricing page sorts by; cost-per-outcome is what your monthly invoice reflects. They diverge exactly where the work is hardest.

This does not generalize downward. On short, simple, high-volume tasks, you're paying double for a capability ceiling you never touch, and there Opus 4.8's economics are clearly better.

Where each one earns its place

Reach for Fable 5 when:

• You're refactoring or migrating a large codebase, where coherence has to hold across hundreds of files and thousands of steps.
• You're building a complex app or an autonomous agent, where planning quality and staying on track over a long horizon determine whether it finishes at all.
• The task is genuinely at the frontier and a higher pass rate is worth paying for, because a cheaper model failing the task costs more than an expensive one completing it.
• You need the strongest context handling across very large inputs.

Stay on Opus 4.8 when:

• You're doing daily coding, bug fixes, small features, quick scripts, where both models land in the same place and the cheaper one wins on value.
• Volume is high and margins are thin; half the token price compounds fast at scale.
• You want maximum predictability. This is the underrated one: Fable 5 carries safety classifiers for cybersecurity, biology, chemistry, and related topics, and routes flagged prompts to Opus 4.8 mid-session. Anthropic reports this happens in under 5% of sessions; independent testing put it higher on security-heavy work. It's tuned conservatively, so it occasionally catches harmless requests. If your work brushes those domains, Opus 4.8 gives you a consistent responder with no fallbacks.
• Stylistic consistency matters to you. Some users find Opus 4.8 has steadier "taste" across a long project, a subjective call, but a real one.

What using them actually feels like

Benchmarks describe the ceiling; daily use fills in the texture. The consistent reports from developers:

Fable 5 is noticeably stronger at planning and executing long multi-step projects. It holds a thread across a large codebase or a long conversation where a weaker model drifts, and it often reaches the finish in fewer total turns. The upgrade feels less like "a bit smarter" and more like a genuine version bump on hard problems.

Opus 4.8 feels more like a dependable daily driver: fast, cheaper, no interruptions, and on ordinary tasks you'd struggle to tell the two apart in a blind test. The gap that looks dramatic on a leaderboard shrinks to almost nothing on a normal Tuesday's worth of coding.

The honest summary from the community: the difference is very real on hard agentic work and easy to overstate everywhere else.

A note on availability

Fable 5's first month was eventful. It launched June 9, was suspended June 12 to comply with a U.S. Department of Commerce export-control order, and was restored on July 1, 2026 after the order was lifted on June 30, now available across Claude.ai, the Claude Platform, Claude Code, and Cowork. (Mythos 5, the same underlying model with the safety classifiers lifted, remains limited to approved partners.) As of this writing, both models are fully available; if you're reading this later, it's worth confirming current status, since this space moves quickly.

The bottom line

Don't think of this as choosing a model. Think of it as building a routing rule.

Send your hard, long, autonomous work to Fable 5: the frontier coding, the big migrations, the agents that have to plan, because there its capability advantage is decisive and, on the hardest tasks, it can even be cheaper per completed job. Keep everything else on Opus 4.8: the daily coding, the high-volume work, the cost-sensitive tasks, and anything that brushes the domains where Fable 5's classifiers would reroute you anyway. Half the price, none of the fallbacks, and on ordinary work you won't feel the difference.

The teams getting the most out of this generation aren't the ones who picked a side. They're the ones who matched the model to the task, paying for the ceiling only when they actually need to reach it.