Fable 5 vs Sonnet 5: what a 3x token price gap actually costs you

The short answer

Claude Fable 5 (launched June 9, 2026) is Anthropic's most capable generally available model, priced at $10 per million input tokens and $50 per million output tokens. Sonnet 5 (launched June 30) runs at an introductory $2/$10 through August 31, then its standard $3/$15.

On list price, Fable 5 is ~3.3–5x Sonnet 5 depending on the window. That gap is real and shows up even on trivial prompts — but it isn't the whole story.

The "Hello" test and why it's misleading

A cost comparison that circulated widely found a single "Hello" prompt cost over 3x more on Fable 5 than Sonnet 5 ($0.47 vs ~$0.15). The headline is true to the tariff, but the per-task framing obscures what actually drives cost:

• Output tokens dominate. Fable 5 returned ~80 output tokens on "Hello" versus Sonnet's 141 — so per token Fable was more expensive, but raw prompt cost was dominated by the session's context, not the one-word reply.
• Cache behavior shifts the real number. Sonnet read far more tokens from cache (cached reads are ~90% cheaper), which is why its bill was lower even with more output. Whoever caches better wins, model-to-model.
• A single "Hello" is a bad benchmark. It measures the tariff, not the workload. The ratio matters, not the absolute cents.

Where the cost gap actually shows

On real coding tasks the pattern holds but gets messy, because it's not just the sticker price:

• Output volume. The high-end model tends to write more (longer comments, more explanation). More output at 5x output price compounds fast.
• Slow starts / weaker caching. If the big model re-reads context instead of hitting cache, every subsequent turn is more expensive.
• Agent loops. In Claude Code, one "task" is many model calls. A 3x per-token gap becomes a 3x gap on every tool call in the loop.

The practical rule: model tier dominates prompt tuning as a cost lever. You can't save a 5x price gap by writing a better prompt.

The tokenizer gotcha

Sonnet 5 shipped an updated tokenizer that can map the same input to roughly 1.0–1.35x as many tokens, depending on content. This matters when migrating between models or comparing two models' token counts — identical prompts can bill different input token totals. Any cost comparison that doesn't control for this is comparing apples to oranges.

When to pay for Fable-class

Fable 5's case is narrow and expensive:

• The hardest autonomous knowledge work / long-horizon coding, where a single Fable run replaces many cheaper runs.
• Problems where 50-million-line codebases or frontier reasoning genuinely need the extra capability.
• When token efficiency offsets the price: Anthropic reports Fable 5 doing in a day what took its team over two months by hand, using a third of the reasoning tokens on frontier physics research.

For everyday coding — feature work, refactors, tests, debugging — Sonnet-tier is the price/performance default. That matches evidence that Opus 5 deliberately launches at half Fable 5's price, positioned as "Fable-5-adjacent at ~half the cost."

The one habit worth stealing

Don't trust anyone's benchmark — including this one. The reliable move is a fresh-session cost check on your own workload:

1. Pick 3–5 representative tasks.
2. Run each in a fresh session per model (so cache starts cold).
3. Record session cost, output tokens, and cache reads/writes.
4. Compare per-task cost, not per-token price.

Cold-cache fresh sessions matter because the moment two models have different cache histories, their numbers stop being comparable.

Takeaway

Sonnet 5 is the sensible default for most Claude Code work; Fable 5 is reserved for the hardest, most autonomous tasks where its token efficiency can pay for the tariff. Budget-conscious teams should treat model tier, not prompt quality, as the primary cost lever — and measure fresh-session cost on their own tasks before deciding.

---

Primary sources: Claude Fable pricing and the Anthropic pricing page. Model pricing and launch dates verified against Anthropic sources on 2026-08-07; the article's per-session cost figures are the secondary author's measurements, not Anthropic's.