Claude Fable 5 vs Sonnet 5: Which Model for Which Job

Use Sonnet 5 for 90% of your work and escalate to Fable 5 only when the problem is genuinely hard — that’s the routing rule, and it exists because Fable 5 is metered by usage credits even on paid plans (as of July 7, 2026) while Sonnet 5 is the included default. Model choice is now a budgeting decision, not just a quality one. Opus 5, which landed on July 24, 2026, adds a middle rung to that ladder.

The lineup, plainly

Model	What it is	Cost signal
Sonnet 5	The workhorse. Default on Free and Pro.	Included; API $2/M in, $10/M out — permanent since Aug 11, 2026
Opus 5	The value rung. Close to Fable 5’s capability, 1M-token context. Default on Max.	API $5/M in, $25/M out — half Fable 5’s rate
Fable 5	The frontier. First of the Claude 5 family, above Opus.	Usage credits on plans; API $10/M in, $50/M out
Haiku 4.5	Fast and cheap.	High-volume API work

Jobs for Sonnet 5 (the default)

Drafting, editing, summarizing, everyday coding, brainstorming, email, scripts, research synthesis. If you’re unsure which model a task needs, it needs Sonnet 5 — the quality bar of “default” moved up a full generation this year, and most operators haven’t recalibrated.

Jobs for Fable 5 (the escalation)

• Multi-constraint reasoning — legal review, architecture decisions, anything where missing one interaction between requirements is the failure mode
• Hard debugging — the bug Sonnet went in circles on
• High-stakes single documents — the proposal or analysis where a 10% quality gain has dollar value
• Long agentic runs — complex multi-step work where a smarter plan saves more credits than it costs

The tell that you should escalate: you’ve re-prompted Sonnet 5 three times on the same problem. Three retries on the workhorse costs more attention than one clean pass on the frontier model.

Where Opus 5 fits (as of July 24, 2026)

Opus 5 is the rung between the two. Anthropic’s own framing is near-Fable-5 capability at half the API price, with a 1M-token context window in and 128K out. Two jobs it takes cleanly:

• API work at volume where Fable 5’s $10/$50 doesn’t survive the spreadsheet but Sonnet 5 keeps missing. $5/$25 is the middle price for the middle problem.
• Whole-repository or whole-archive passes where the constraint is how much you can put in the window, not how hard the reasoning is.

On plans, the practical rule is simpler: Opus 5 is the default on Max and the strongest model you can reach on Pro, so if you’re on Max the escalation is already made for you. Fable 5 stays the ceiling; Opus 5 is what you reach for before paying ceiling prices.

The credit discipline

Fable 5’s metering means an undisciplined “always use the best” habit gets expensive. Run it like a specialist on retainer, not a daily hire:

1. Start every task on Sonnet 5.
2. Escalate on evidence (the three-retry tell), not on vibes.
3. Escalate one rung at a time — Sonnet 5, then Opus 5, then Fable 5. Most tasks that beat Sonnet stop at Opus.
4. For API work, cache aggressively — the 90% input-token discount on cached prompts changes Fable 5’s math entirely.

Tools need jobs. Models do too. Sonnet 5 has the job called “everything”; Opus 5 has the job called “everything, but bigger and harder”; Fable 5 has the job called “the thing everything couldn’t do.”