The Claude Fable 5 Playbook
Anthropic's most capable model is included in Claude subscriptions through July 7 — then it moves to usage credits. Here's how operators extract maximum value from the window: what to run it on, how to prompt it, and what to bank so the gains outlast the model access.
Window closes July 7 — after that, Fable 5 bills via usage credits until Anthropic restores subscription capacity.
~14 min read
A new tier above Opus — not an incremental bump.
Fable 5 is the first model in Anthropic's Claude 5 family, in a new Mythos-class tier that sits above Claude Opus. It's built for the work previous models couldn't finish: long autonomous runs, hard architecture calls, adversarial review, and research synthesis.
Tier
Mythos-class — above Opus 4.8, Anthropic's most capable released model
Context window
1M tokens (entire codebases, full document sets in one session)
Max output
128K tokens per request
Reasoning
Thinking is always on — it deliberates before every answer
Turn length
Single responses on hard tasks can run many minutes — by design
API pricing
$10 / $50 per million tokens in / out (claude-fable-5)
How we got a window
July 1
Fable 5 returns to Claude after the U.S. lifted export-control restrictions on Anthropic's most powerful models — available on Claude.ai, Claude Code, and the API.
Through July 7
Included for Pro, Max, Team, and select Enterprise plans at up to 50% of weekly usage limits. This is the window.
After July 7
Available via usage credits. Anthropic says it intends to restore Fable to subscriptions once capacity allows — but no date is committed.
Spend it on judgment, not keystrokes.
The gap between Fable 5 and everything else shows up in wide-solution-space work — decisions, reviews, research, architecture. It barely shows up in mechanical edits any model does fine. So invert your queue: pull forward everything that's hard to think about, defer everything that's just typing.
Worth the window
- Architecture and data-model decisions you've been deferring
- Adversarial review of code you're about to ship — security, billing, auth
- Red-teaming your own product, pricing, or launch plan
- Deep research that ends in a decision — market, vendor, vertical
- Your hardest unsolved problem — the one you've restarted three times
- Writing the specs, rubrics, and prompts your team will reuse
Don't burn it on
- Boilerplate, CRUD endpoints, and formatting cleanups
- Long mixed sessions that drift across six unrelated tasks
- Work any model completes identically — renames, config edits
- Vague prompts you'd have to clarify three times anyway
- Automations that will still be running after the window closes
- Anything you won't write down — unbanked output evaporates
Six plays, ranked by leverage.
Run these in order. Each one converts temporary model access into a durable asset — a decision, a finding, or a spec that keeps paying after July 7.
Feed it your hardest unsolved problem first
The teams that got the most out of early access gave Fable their hardest open problem on day one — the migration that stalled, the bug nobody can reproduce, the system design with no obvious shape. Have it scope the problem, ask its questions, then execute. If it only gets 80% of the way, that 80% is a map you didn't have.
Make the decisions you've been deferring
Pricing model, schema design, build-vs-buy, which market to enter. These are judgment calls with wide solution spaces — exactly where the tier gap is largest. Force a decision with written reasoning: options considered, tradeoffs, the call, and what would change it. A decision record outlives any model window.
Audit everything you're about to ship
Run deep review passes on the code paths that can hurt you: payments, auth, data handling, anything customer-facing. Ask for every finding with confidence and severity — filter later, not during. Findings written down now can be fixed by any model (or any engineer) next month.
Red-team your own product and plan
Ask it to refute you: "Find the case where this guardrail fails." "Argue this launch misses." "Break this pricing." Refutation quality is where frontier models separate most clearly from the tier below — and a survived objection is worth more than ten confirmations.
Run research that ends in a report
Vendor selection, competitive landscape, regulatory exposure, market sizing. Demand citations and an explicit recommendation, not a survey. The deliverable is a document with sources your team can check — synthesis this good is normally a consultant invoice.
Lock your specs, prompts, and skills
Have it write and pressure-test the artifacts that direct future AI work: API contracts, system prompts, evaluation rubrics, reusable skills. This is compounding leverage — every model you use afterward inherits the quality of the specs Fable wrote.
It's not a faster Opus. Drive it differently.
Fable 5 rewards a different working style. Eight adjustments that change output quality more than any clever phrasing:
Full spec up front, one turn
Give the whole task — goal, constraints, context, what done looks like — in the first message. It plans across the entire horizon; drip-feeding requirements wastes its planning on stale information.
State goals, not steps
Step-by-step scaffolding written for older models actively reduces Fable's output quality. Say what and why; let it decide how. If you have a prompt library, A/B it with the hand-holding removed.
Give the reason behind the request
"I'm deciding X for customer Y, they need Z — with that in mind: [request]." It connects intent to detail instead of guessing at it, and the output lands closer on the first pass.
Ask for refutation, not validation
"Try to break this." "Argue the opposite." "What would make this wrong?" You get its strongest reasoning when you point it against your own position.
One hard problem per session
Its edge compounds over a long, coherent context on a single problem. A session that wanders across six tasks spends that coherence on noise. Start fresh for each big swing.
Grant autonomy on the small stuff
Tell it to make minor calls itself — naming, defaults, equivalent approaches — and only check in on scope changes or destructive actions. Otherwise you'll trade its long-horizon strength for a permission ping-pong.
Give it somewhere to write things down
A plain notes file it can read and update makes long engagements measurably better — decisions, dead ends, and context persist across sessions instead of evaporating with each one.
Expect long turns — that's the feature
A hard request can run many minutes while it gathers context, builds, and verifies its own work. Don't interrupt it to check progress; queue your next problem instead.
Bank artifacts, not vibes.
The test for every Fable session: what document exists afterward? Model access is temporary; what you write down is permanent.
Compounds after July 7
- Decision records — the call, the reasoning, what would reverse it
- Audit findings with severity, file references, and repro steps
- Specs and API contracts any model or engineer can implement against
- Research reports with citations your team can verify
- Prompts, rubrics, and skills that direct future AI work
Evaporates or backfires
- Half-finished code a weaker model then has to interpret
- Conclusions that live only in a chat transcript
- Automations scheduled to run after the 7th that assume Fable-level judgment at runtime
- "We'll remember the reasoning" — you won't; write it down
Fable 5 FAQ
What exactly is Claude Fable 5?
The first model in Anthropic's Claude 5 family and their most capable released model — a new Mythos-class tier above Claude Opus. It shares its underlying model with Claude Mythos 5, which is restricted to approved organizations; Fable 5 is the generally available version with additional safety measures.
What happens after July 7?
Fable 5 moves from being included in Pro, Max, and Team subscriptions to billing via usage credits. Anthropic has said it intends to bring it back into subscriptions once there's enough capacity, but hasn't committed to a date. It remains available on the API ($10/$50 per million tokens).
Is it worth paying usage credits for after the window?
For judgment-heavy work — architecture, audits, research, decisions — usually yes; the output quality gap justifies it. For routine coding and drafting, Opus 4.8 or Sonnet remain the economical default. The playbook above is exactly the work worth paying for.
How is it different from Claude Opus 4.8 in practice?
Longer coherent autonomous runs, stronger adversarial reasoning, better first-shot implementations of well-specified systems, and better synthesis across large document sets. On short mechanical tasks the difference is small — which is why this guide says to spend the window on hard problems.
Why did it disappear and come back?
Fable 5's initial availability was interrupted by a U.S. export-control restriction on Anthropic's most powerful models. That restriction was lifted, and the model was redeployed on July 1, 2026 with strengthened safeguards and the temporary subscription window through July 7.
Four days is enough — if someone's done this before.
This playbook is what we run for clients: frontier-model audits, decision sprints, and AI systems installed into real operations. If you'd rather ship it than study it, get on a call before the window closes.