The 10/80/10 Rule: How to Use Fable 5 Without Burning Through Your Budget

What Fable 5 is

Fable 5 is Anthropic's most capable publicly available model. It sits in the Mythos family and costs roughly twice what Opus 4.8 costs per token.

The price gap is real, but the intelligence gap is also real: Fable 5 is meaningfully better at architecture decisions, catching subtle logic errors, and reasoning through edge cases that Opus 4.8 would miss. The question isn't whether to use Fable — it's whether to use it for everything.

The 10/80/10 rule

Most development work breaks into three phases. Map a model to each based on where higher reasoning actually changes the outcome:

First 10% — plan in Fable. Architecture decisions, data models, identifying edge cases, writing the spec. This is where wrong decisions compound. Fable's reasoning advantage is worth the cost here.

Middle 80% — build in Opus 4.8. Components, boilerplate, boring refactors, implementations that follow a clear spec. Opus 4.8 handles this well at half the cost. Fable's extra reasoning doesn't improve output quality on predictable execution work.

Last 10% — review in Fable. The final pass for nasty bugs, logic errors, and security vulnerabilities. Fable also blocks in high-risk areas — cyber, bio, chemistry — and falls back to Opus 4.8 by design.

The tagline: plan in Fable, build in Opus, review in Fable. You're paying for its brain, not its typing.

Where to apply each model

Fable 5:

• System architecture and data model design
• Identifying edge cases and failure modes before writing code
• Security reviews and final logic checks
• Any decision that's hard to reverse once built

Opus 4.8:

• Writing components against a clear spec
• Refactoring and cleanup
• First drafts, research, quick lookups
• Creative work and execution-oriented tasks where the decision is already made

The boundary: if the wrong answer here costs more than the token price difference, use Fable.

The Mythos distinction

Fable 5 is the public version. "Mythos" is the locked-down partner-only variant — same family, higher capability ceiling, available only through specific Anthropic partnerships. If you're accessing models through Claude.ai, Claude Code, or the standard API, you're using Fable 5. Mythos is not available through the standard tier.

Why Opus 4.8 is still the daily driver

The 10/80/10 rule means Fable 5 handles 20% of your work at most. For everything else — creative work, first drafts, quick research, execution tasks — Opus 4.8 is faster and cheaper with no meaningful quality gap on those task types.

Running everything through Fable isn't a strategy, it's waste. The routing decision is the strategy.

Frequently asked questions

How do I switch between Fable 5 and Opus 4.8 in Claude Code? Use the model selector in the session header. You can set a default in settings and override per-session. In the API, pass

model: 'claude-fable-5'

model: 'claude-opus-4-8'

Is the 10/80/10 split exactly right for every project? No — it's a starting heuristic. Some projects are almost entirely architectural (Fable-heavy); others are mostly boilerplate (Opus-heavy). The principle is the useful part: identify which decisions are high-stakes and hard to reverse, apply the expensive model there.

Does this routing work in Claude Code or only custom API setups? Both. In Claude Code, switch models manually between phases or use a routing skill. In API setups, classify the task type and select the model in code. The principle is the same: match model capability to task complexity.