Claude Fable 5 is probably the strongest model you can publicly get your hands on today. It is also how one builder burned $371 in half a day.

Both facts matter, because the right way to use this model is not as your daily driver.

Under the hood, Fable 5 is essentially Claude Mythos with public-facing safety guardrails bolted on. Mythos only goes to approved organizations. Fable is the version the rest of us get.

The benchmarks that actually matter

Fable 5 pulls clearly ahead on SWE-bench Pro and FrontierCode, two benchmarks that map closer to real engineering pain than most: complex code repair, repo-scale changes, long reasoning chains, and whether the final diff is actually mergeable.

  • SWE-bench Pro (complex code repair): Fable 80.0 vs GPT-5.5 58.6. A 21.4 point gap.
  • FrontierCode Diamond (frontier coding, xhigh difficulty): Fable 29.3 vs GPT-5.5 5.7. That is 5.1x.
  • Terminal-Bench 2.1 (terminal agent tasks): Fable 84.3 vs GPT-5.5 83.4. Basically a tie.

The pattern is clear. Fable's real edge shows up in long-horizon, cross-module engineering work. On simple tasks, most frontier models feel about the same.

Pricing

Rough math: Fable runs about 1.7-2x the cost of GPT-5.5 depending on your token mix, and about 2x the cost of Opus 4.8.

The good news: Claude Pro, Max, Team, and per-seat Enterprise users get Fable 5 free until June 22. After that it switches to usage-based pricing, and the bill climbs fast.

Notes from real use

1. It is genuinely expensive

Community reports back this up. One Max 20x subscriber burned through a full 5-hour window quota in 45 minutes. Another posted that $371 half-day bill, and that was with roughly 92% cache hits.

2. Execution and skill-following are excellent

I have a Word formatting skill that got flaky after I migrated from Windows to Mac. Same model, same agent, but every run came out different, like it had lost part of its memory. I switched to Fable 5 and it nailed the job on the first run.

3. It fixes the bugs other models give up on

A project crash that Opus 4.8 had been chewing on for over a week, Fable solved in 10 minutes.

4. Frontend and UI work got noticeably better

I fed a GPT Image design mockup into Claude Design with Fable behind it. It extracted the art assets on its own and produced a more sensible implementation than Opus managed on the same job.

5. The safety filter is too aggressive

The classifier is jumpy. Some topics silently fall back to Opus 4.8 mid-conversation. Biomedical questions get downgraded to Opus 4.8 100% of the time.

What the leaked system prompt tells developers

Fable 5's system prompt has leaked. A few details worth knowing if you build with it:

  • It is told to ask fewer questions: act on ambiguous requests, reach for tools when possible, create files when possible.
  • Before producing docs, slides, PDFs, spreadsheets, or code, it must read the relevant SKILL.md first. Those files carry environment constraints, rendering pitfalls, output paths, and library limits. Skipping them degrades output quality.
  • Anything over 10 lines of code goes into a file, not the chat.
  • Outputs cannot be used to improve competing models or to train frontier LLMs.

Full system prompt: https://gist.github.com/hqman/fb7b5caa77f310b6c9016a26c7cc2cd7

Bottom line

Route by difficulty.

Send the hard, multi-module, long-horizon work to Fable 5. Keep the everyday tasks on Opus 4.8 or GPT-5.5, where the economics still make sense.