Claude Fable 5 is Anthropic's most capable widely released model, positioned above Opus 5 for the hardest reasoning problems and the longest autonomous agentic runs. It is not the model you reach for by default: it costs more than Opus 5 ($10 input / $50 output per million tokens versus Opus 5's $5/$25), thinking cannot be turned off, and a single request on a hard task can legitimately run for many minutes. Claude Mythos 5 is effectively the same model under a different name, available exclusively to organizations in Anthropic's Project Glasswing program — everyone else uses Fable 5 directly. The honest framing: Fable 5 is what you reach for when Opus 5 is capable but not quite reliable enough on a specific, genuinely hard workload, not a universal upgrade.
How it works
Fable 5 shares its overall API shape with Opus 5 — the same 1M-token context window, which is also the maximum rather than just the default, the same 128K max output, and the same effort parameter for controlling reasoning depth. The differences are deliberate constraints layered on top:
- Thinking is always on. There is no way to disable it. Omitting the thinking parameter, or setting it to adaptive explicitly, both work; explicitly disabling it returns an error. Depth is controlled entirely through the effort parameter instead.
- The raw chain of thought is never returned. You get either an empty thinking field by default or a readable summary if you opt in — never the unfiltered reasoning trace. This is a hard privacy and IP boundary, not a UI setting.
- It needs 30-day data retention. Requests from organizations configured for zero data retention are rejected outright, so this is worth checking on your account before you build against it.
- Safety classifiers target research biology and cybersecurity content specifically. Benign, adjacent work in security tooling or life sciences can occasionally trip a decline. Anthropic recommends configuring an automatic fallback to Opus 5 for exactly this reason.
What Fable 5 is actually good at
| Strength |
What it looks like in practice |
| Long-horizon autonomous work |
Multi-hour agent runs that gather context, build, and self-verify without losing the thread |
| First-shot implementation |
Turning a well-specified system design into a working implementation in one pass |
| Enterprise deliverables |
Multi-sheet spreadsheets with real formulas, polished slide decks, long structured documents |
| Dense or degraded vision |
Reading flipped, blurry, or noisy images by cropping and re-analyzing rather than guessing |
| Parallel sub-agent coordination |
Delegating independent workstreams and reliably tracking what each one reports back |
The gap over Opus 5 is smallest on short, well-scoped tasks — a quick bug fix or a single-file edit rarely benefits from Fable 5's extra reasoning, and the added latency is not worth it there.
Common mistakes
- Defaulting to Fable 5 because it's "the best model." It is priced and built for a narrow set of hard problems. Using it for routine coding or chat traffic is expensive and slower for no real quality gain.
- Not planning for long turns. A Fable 5 request on a genuinely hard task can run well past what a typical HTTP client timeout expects. Stream the response, and design your product around asynchronous progress rather than a blocking call.
- Forgetting the data retention requirement. Teams on zero data retention sometimes discover Fable 5 rejects every request only after they've built against it — verify your organization's retention settings first.
- Not handling declines. Treat a refusal as a normal response to branch on, not an exception. Pair Fable 5 with an automatic fallback to Opus 5 so a benign security- or biology-adjacent prompt doesn't just dead-end.
FAQ
Is Claude Fable 5 the same as Claude Mythos 5?
Functionally, yes — same capabilities, pricing, and API behavior. Mythos 5 is the name used exclusively inside Anthropic's Project Glasswing program; everyone else calls the model Fable 5.
Why is Fable 5 priced above Opus 5?
It targets a narrower, harder slice of work — always-on maximal reasoning and the longest agentic runs — rather than general-purpose production traffic, and Anthropic prices it accordingly.
Can I see the model's actual reasoning?
No. You can request a readable summary of the thinking, but the raw chain of thought is never exposed on this model, unlike some server-side tool outputs elsewhere in the API.
Should I use Fable 5 for a customer-facing chatbot?
Almost never. The latency and cost profile fit long-running, high-stakes background work far better than an interactive chat turn expecting a fast reply.
Where to go next
For a hands-on walkthrough of when and how to actually put Fable 5 to work, see our practical guide to Fable 5 use cases. If you're still deciding whether you need this tier at all, compare it against the flagship in our Claude Opus 5 review, and see the full cost picture in Claude 5 family pricing explained.