Fable 5 use cases — what it's for, with receipts.

Six receipts shipped with the announcement. Here's each one, with the read Anthropic's copy won't give you. Ch 24's rule applies to customer quotes the same as it applies to benchmark rows: a receipt is a signal, a private eval is the proof.

1. Stripe — 50 million lines of Ruby, migrated in one day. Scoped at two months by hand.

   The read: the strongest cost-per-task receipt any model launch has shipped — not a demo, a migration with a by-hand baseline attached. Also: one company, one codebase, vendor-reported. The Ch 29 math on this lives on the pricing page.

2. Cursor + Cognition — the coding-tool vendors agree. Cursor calls Fable 5 the "state of the art model on CursorBench". On Cognition's FrontierCode, it scores highest among frontier models at medium effort.

   The read: two companies whose product economics depend on picking the best model picked the same one. "At medium effort" is the detail worth keeping — the top score didn't need the max dial.

3. GitHub — the long-horizon quote. Praised complex, long-horizon coding tasks with "a level of autonomy and reliability that exceeded previous benchmarks".

   The read: "long-horizon" is the load-bearing phrase. It matches the Claude Code banner positioning, and it's the axis where the table gap is widest — FrontierCode Diamond (xhigh effort) at 29.3% vs Opus 4.8's 13.4%.

4. IMC — the trading desk. Fable 5 "aced their trading-analysis evaluations nearly across the board".

   The read: the first non-coding receipt, and the most instructive one — IMC ran private evals, which is the Ch 25 discipline. "Nearly" is the honest word in the quote; keep it.

5. Drug design — around 10x. Anthropic-internal protein experts accelerated aspects of the drug-design process by around 10x.

   The read: the most caveated receipt on the page — internal experts, "aspects of", "around". Anthropic wrote those hedges themselves. A real signal for science workloads, not a number to put in a deck.

6. Genomics — autonomous research, but read the model name. Mythos 5 conducted autonomous research identifying cell roles across 138 animal species and trained a custom ML model 100x smaller than a recent Science-journal publication's. Scientists preferred its hypotheses ~80% of the time in blinded comparisons against Opus-class models.

   The read: the most impressive receipt of the launch — and it isn't Fable 5.

Benchmarks can be reward-hacked; a timelapse anyone can watch end to end can't. Anthropic shipped two with the launch, and they're the most honest evidence on this page — you can verify them with your own eyes.

Pokémon FireRed, vision only. Previous Claude models couldn't finish FireRed even with harnesses handing them maps and navigation tools. Fable 5 beat the game on raw screenshots — no maps, no navigation aids, vision only. That's the long-horizon claim in its purest form: an entire playthrough's game state held across a single run with the thinnest possible harness.

The solar system, simulated. The second demo has Fable 5 simulate the solar system and predict a solar eclipse — the science-workload counterpart to the genomics receipts above, except this one you can watch.

Every receipt above maps onto a pattern this book already documents. That's the useful translation: not "Stripe did a cool thing" but "which of my workloads has the Stripe shape".

1. 1 — Long-horizon refactors and migrations. The Stripe shape. This book ran the same pattern one size down: Ch 43's run deleted a net 91,874 lines across 718 files behind layered verification. Stripe's day is that move at 50M-line scale. If you have a migration scoped in months, this is the workload Fable 5 was priced for.
2. 2 — Run-until-done loops. "Complex, long-running work" is Anthropic's own banner copy, and it is the Ch 38 pattern verbatim: goal in, evaluator on, agent loops until the condition proves. A model that holds the thread longer is worth more per loop, not per token.
3. 3 — Swarm orchestration — the conductor seat. Put the expensive model where the judgment is and fan cheap ones out below it (/swarms, /dynamic-workflows). Fable 5 takes the conductor seat; Haiku and Sonnet stay in the sections. The 2× sticker buys one seat, not the whole orchestra.
4. 4 — Knowledge work — the quiet headline. GDPval-AA reads 1932 vs Opus 4.8's 1890. GDP.pdf — vision knowledge work, no tools — reads 29.8% vs 22.5%. The launch is sold on code; the table says the gain shows up in contracts, decks, and documents too. If you're a non-developer reading this, these two rows are yours.
5. 5 — Computer use — eval before believing. OSWorld-Verified: 85.0% vs Opus 4.8's 83.4%. An incremental gain, not a leap — if browser-driving agents are your workload, the delta may not survive a 2× price. Measure it.

Where this runs day to day for most operators: Fable 5 in Claude Code — the banner, /model, and the routing call.

• Cyber- and bio-adjacent work. The safeguard classifiers trip — in the Claude apps the response comes from Opus 4.8 instead; on the API the request errors, unbilled, unless you opted into Opus-priced fallback. Either way you bought Fable and got Opus. Route those workloads to Opus 4.8 directly.
• Bulk cheap work. Classification, extraction, routine drafts — Haiku 4.5 at $1/$5 and Sonnet 4.6 at $3/$15 didn't get worse on June 9. The tier list stays the live source of truth on routing.
• Anything you haven't evaled. Fable 5 is included in paid-plan limits June 9–22, 2026 — a free private-eval window (Ch 25). After June 22 the meter runs on usage credits. Eval first, route second.

The receipts say what Fable 5 did for six teams that aren't yours. The free window says what it does for you. One of those is marketing.