Yes, your existing skills work with Claude Fable 5, unchanged. A skill is a SKILL.md folder in Anthropic's open format, and the format is model-agnostic, so nothing you built for earlier Claude models breaks. What changes is not compatibility. It is the stakes: Fable 5 works autonomously for far longer, so a sloppy skill now runs off the rails for longer too.
I build skill tooling for a living, so when a new flagship model lands the first question I get is some version of "do I have to redo all my skills?" The short answer is no. The longer answer, the one worth 1,200 words, is that a more capable model quietly moves where your attention should go. Let me walk through what actually changed when Fable 5 launched on June 9, 2026, what the suspension in the middle did and did not do, and how to think about the model your skills run on.
Do existing Claude skills work with Fable 5?
They do, and this is the least surprising thing about the launch. If you have read our primer on what a Claude skill is, you already know a skill is just a folder with a SKILL.md file: a name, a description, and plain-English instructions underneath. There is nothing model-specific in that file. It never names a model, never pins a version, never assumes anything about the engine reading it.
So when Fable 5 arrived, it read the same skill folders that Sonnet, Opus, and every prior model read. The same is true the other direction. A skill you write today for Fable 5 still runs on Sonnet 5, still runs in the Claude apps, still runs in Codex and Cursor and Gemini CLI. That portability is the whole point of the open format, and it is exactly why a model upgrade is a non-event for compatibility. You do not update your skills for Fable 5 any more than you rewrote your onboarding docs when you hired a smarter person.
What actually changed for skills when Fable 5 launched?
The interesting change is capability, not format. Fable 5 and its restricted sibling "can work autonomously for longer than any previous Claude models", with a 1-million-token context window and up to 128,000 output tokens per request. Anthropic points to a Stripe migration that Fable finished in a day when it would have taken a team over two months. That is the headline: sustained, multi-stage, self-verifying work with far less human check-in.
For a skill, that cuts two ways. A good skill compounds. Instructions that used to guide a model through three careful steps now guide it through thirty, and the model holds the thread across all of them. But a skill with a buried mistake compounds the same way. On an older model that stopped and asked every few steps, you caught the drift. On a long-horizon model that keeps going, the same flawed instruction gets applied thirty times before anyone looks. The skill did not get worse. The blast radius did.
Why does skill quality matter more on a long-horizon model?
Here is the sentence I keep repeating to people: a model that works autonomously for days will execute a bad skill for days. Autonomy is a multiplier, and it multiplies whatever you feed it.
Think about a skill that cleans up invoices and has one loose rule: "use the invoice date." On a short-horizon model, you review each batch, spot the invoices with no date, and fix them by hand. On Fable 5 running unattended, that ambiguity does not surface as a question. The model resolves it however it resolves it, applies that choice to the whole pile, and hands you a finished result that looks polished and is quietly wrong in a hundred places. The failure is not louder. It is more consistent, which is worse.
This is why the boring disciplines matter more now, not less. Tight, single-purpose skills. Sharp descriptions so the right skill triggers and the wrong one stays quiet (our piece on how skills load covers the trigger mechanics). And above all, the testing loop we lay out in the Cowork skill development guide: run the skill on real inputs, read the output like a junior hire's first draft, and turn every miss into a written rule. A skill that survived ten real runs and three corrections was always the goal. On a long-horizon model it stops being a nice-to-have.
The genuine safety net, though, is skill versioning. If a skill can now do a week of damage in an afternoon, you want immutable version history and one-click rollback, so the moment you spot the drift you can pin back to the last version you trusted while you fix the new one. That rollback is the difference between "we caught a bad skill" and "we are unwinding three days of automated output." Purpose-built skills like skill-writer and writing-skills help the agent draft and self-check skills before they ship, which pushes the catch earlier, where it is cheap.
What happened during the Fable 5 suspension, and did skills break?
This is where July got strange. On June 12, the U.S. Commerce Department imposed export controls on Fable 5 (and the restricted Mythos 5) after Amazon researchers found a jailbreak, and the order forced Anthropic to cut the models off for any foreign national. Unable to verify nationality in real time, Anthropic disabled both models globally. For about three weeks, you could not select Fable 5 at all.
Did skills break? No, and this is the model-agnostic point made concrete. During the suspension, your skills simply ran on whatever model you fell back to (Opus 4.8, Sonnet, Haiku), because the SKILL.md folder does not care which engine reads it. Anthropic even reset the 5-hour and weekly rate limits to make the fallback models more usable while Fable was dark. The skills that ran on Fable on June 11 ran on Opus on June 13 with no edits. That is the compatibility guarantee doing exactly what it promises, under stress.
Fable 5 came back on July 1 across Claude.ai, the Claude Platform, Claude Code, and Cowork, after the export controls were lifted on June 30. It returned with a new safety filter aimed at that specific jailbreak, one that blocks the attack in more than 99% of tries and routes flagged requests to the weaker Opus 4.8, notifying you when it does. Worth knowing: Anthropic says that trade-off raises false alarms on routine coding tasks. So a skill that touches security-adjacent work may occasionally get silently downgraded to Opus 4.8 mid-run. It will not error. It will just be handled by a less capable model for that request.
Which model should you run skills on, cost-wise?
This is the practical question, and as of early July 2026 the pricing gap is real. Fable 5 runs $10 per million input tokens and $50 per million output tokens, double the Opus 4.8 rate. Sonnet 5, meanwhile, is on an introductory promo of $2 input and $10 output through August 31, 2026, rising to $3/$15 after. That is a 5x spread on output tokens.
For the vast majority of skill-driven work (drafting the weekly report, cleaning the spreadsheet, turning a transcript into minutes), Sonnet 5 runs the identical skill folder at a fraction of the cost, and you will not feel the difference in the output. Reserve Fable 5 for the genuinely long-horizon jobs where its sustained autonomy earns the premium: the multi-hour migration, the sprawling refactor, the research task with fifty sub-steps. The skill is the same either way. The bill is not.
One more current-facts note on Fable's access, because it moved twice in three weeks. On restoration, Anthropic included Fable 5 for up to 50% of weekly usage limits on most paid plans through July 7, after which it moves to metered usage credits at that $10/$50 rate (Search Engine Journal has the breakdown). Check the current terms before you point a long-running skill at Fable and watch the credits drain.
If you are building or storing skills that need to move cleanly between all these models, that model-agnostic layer is exactly what Knack exists to give you: one canonical version of each skill, versioned, that runs wherever Claude runs. Browse the marketplace for working examples to start from.
FAQ
Is Claude Fable 5 available again?
Yes. Fable 5 returned on July 1, 2026 across Claude.ai, the Claude Platform, Claude Code, and Claude Cowork, after the U.S. Commerce Department lifted its export controls on June 30. It came back with a new safety filter targeting the jailbreak that triggered the suspension. Mythos 5, the restricted variant, was restored earlier, on June 26, only for roughly 100 vetted U.S. organizations.
Do I need to update SKILL.md files for Fable 5?
No. A SKILL.md file contains no model-specific fields, so the same folder runs on Fable 5, Sonnet 5, Opus 4.8, and every other Claude without edits. If a skill misbehaves on Fable 5, the fix is the same as always (sharpen the instructions or the description), not a model-specific rewrite. The format is deliberately model-agnostic.
Does Fable 5 have usage limits?
Yes, and they shifted quickly. On its July 1 return, Anthropic included Fable 5 for up to 50% of weekly usage limits on most paid plans through July 7, then moved it to metered usage credits at $10 per million input tokens and $50 per million output tokens. As of early July 2026 those terms are recent, so confirm the current plan details before running long jobs on it.
Is Fable 5 in Claude Code and Cowork?
Yes. When Fable 5 was restored on July 1, 2026, availability explicitly included both Claude Code and Claude Cowork, alongside Claude.ai and the Claude Platform. Your skills run identically in both surfaces, so a skill authored in Claude Code runs in Cowork and back again with no conversion.
What is the difference between Fable 5 and Mythos 5?
They share the same underlying model and published specifications. Fable 5 is the public version with safety classifiers that route sensitive requests (cybersecurity, biology, chemistry) to Opus 4.8 in fewer than 5% of sessions. Mythos 5 has some of those safeguards lifted and is restricted to authorized defenders and researchers, not general users.
Should skills target a specific model?
No, and the format will not let you anyway. The strength of a skill is that it is portable across every model and surface, so pinning it to one engine would throw that away. Choose the model at run time based on the job (Sonnet 5 for everyday tasks, Fable 5 for genuinely long-horizon autonomous work) and keep the skill itself model-neutral.