Claude Fable 5 Is Free Until June 22 — Here's the Play

Last week Anthropic warned that AI is becoming too dangerous. Days later, it shipped Claude Fable 5 — its most powerful model ever, a full tier above Opus 4.8 — to everyone. And if you're on Pro, Max, or Team, you can use it for free until June 22. Both things are true at once, and the contradiction is the most useful signal in the whole release. Here's what a "Mythos-class" model actually changes for the people building on top of it — and why that free window is a deadline, not a gift.

"A tier above Opus" — and what that phrase hides

Fable 5 is the first public model in Anthropic's new Mythos class, a level above the Opus line, and it posts state-of-the-art numbers on nearly every public benchmark — SWE-Bench Verified, SWE-Bench Pro, FrontierCode. It carries the now-standard 1M-token context and up to 128k output per request. Read the benchmark table, nod, move on — because "it's smarter" is the least interesting thing about this launch. Two other facts will shape your roadmap far more than the leaderboard: what it costs, and what it quietly refuses to do.

The price tells you the use case

Fable 5 runs $10 per million input tokens and $50 per million output — roughly double Opus-class on both ends — and it's free for Pro, Max, and Team users only through June 22. That pricing is the positioning. At $50 per million output tokens, this is not your default workhorse. You do not route an agent fleet through it. You reach for it where being right once is worth more than three cheaper attempts: the refactor that touches forty files, the bug that's survived two debugging passes, the architecture call you can't afford to walk back.

A frontier-plus model isn't a better default — it's a scalpel you pick up when the cheap tools have already failed. The teams that route everything through it will get a bill that teaches them the difference. If you don't yet have a rule for which tasks earn the expensive model, that's the gap to close first; I've written the broader version of that discipline in the cost-governance playbook and the per-call token guide.

The gating is the real news

Fable 5 ships with safeguards that re-route a small set of high-risk prompts — cybersecurity, biology, chemistry, distillation — to Opus 4.8 instead of answering directly. And the unguarded version, Claude Mythos 5, isn't public at all: it's locked behind a limited program ("Project Glasswing") for infrastructure providers and vetted cybersecurity researchers.

Sit with what that means for a builder. The capability of the model you call is now conditional. The same endpoint can silently hand you a tier-above-Opus answer on one prompt and an Opus-4.8 answer on the next, based on topic — and you'll pay the Fable rate either way. This is a genuinely new failure mode. You can no longer assume uniform capability across your prompt space. If your product touches security tooling, bio, chem, or anything adjacent, you have to test those paths explicitly, because you may be getting downgraded answers — or refusals — exactly where you expected the most help.

The contradiction is a roadmap, not hypocrisy

"AI is too dangerous" and "here's our most powerful model, free for two weeks" sound contradictory. They're not. They're the shape of the access model that's coming: frontier capability is bifurcating into a safeguarded public tier and a gated raw tier. The most capable version of any model will increasingly be access-controlled, and the version you can actually call will carry guardrails that bite at your edge cases.

Plan for that world now. Stop treating "model capability" as a single number on a benchmark page and start treating it as a spectrum you have to verify — per task, per topic, per tier. The companies that internalize this early will design around it; the ones that don't will ship features that work in the demo and silently degrade in the corner of the prompt space where the safeguards live.

What the free window is actually for

The June 22 deadline is the opportunity. You have until then to learn — for nearly nothing — where a tier-above-Opus model changes your economics. Don't burn it generating marketing copy. Burn it on signal:

• Re-run your hardest, previously-failed tasks. The agentic jobs that Opus or Sonnet botched — feed them to Fable and record what now succeeds. That delta is the only benchmark that matters to you.
• Probe your edge cases. Send the security-, bio-, or chem-adjacent prompts your product actually generates and watch for silent re-routing or refusals. Map where the guardrails sit before a user finds them for you.
• Measure the ratio. For the tasks where Fable wins, is the win worth 2–5x the token cost versus Opus? That number is your routing rule after the free window closes.

What to actually do this week

1. Turn Fable 5 on for your team now — it's free on Pro, Max, and Team through June 22.
2. Pick 3–5 tasks your current models fail or do poorly. Re-run them on Fable and write down exactly what changes.
3. Map your re-routing exposure: test any security/bio/chem-adjacent prompts and note where you get silent downgrades or refusals.
4. After the window, set a hard rule — Fable only for tasks where being right once beats cheaper retries; the default stays on cheaper tiers.
5. Assume the access model keeps splitting. Design your provider layer so swapping or gating a model is a config change, not a refactor.

Anthropic handed you a two-week test drive of the most capable model it has ever released publicly, while telling you out loud that the real thing is too dangerous to ship unguarded. The benchmark scores will be stale in a month. The lesson — that capability is now metered, gated, and conditional — won't be. Spend the window learning your own numbers, not memorizing someone else's.