Fable 5 pricing — the sticker, the window, the real number.

The Anthropic ladder held $5/$25 at the top through Opus 4.7 and 4.8. Fable 5 doubles it — and Mythos 5, the gated twin, carries the same sticker.

One more rung, above the table: Anthropic's own line is that both models cost "less than half the price of Claude Mythos Preview" — which puts the preview's partner pricing north of $20/$100. The withheld model was a premium product for the vetted few; the shipped one costs less than half that and anyone can buy it. The doubling reads different in that frame: the top of the public ladder moved up, the price of frontier access moved down.

Two lanes, different meters. Claude API and consumption-based Enterprise: fully available June 9, metered at $10/$50 from the first token. Subscription plans (Pro, Max, Team, seat-based Enterprise): Fable 5 is included in your existing plan limits from June 9 to June 22, 2026 — then it requires usage credits. Anthropic says it aims to restore Fable 5 as standard on plans when capacity allows, no date given.

Read the window as what it is: two weeks of frontier capacity at zero marginal cost, on the plan you already pay for. That's not a promo, it's a free A/B against your current model — if you run it as one. The meter starts June 22 whether you measured or not.

This site has run this exact play before. On 2026-05-19 Google announced Gemini 3.5 Flash and tripled the token price — $0.5/$3 to $1.5/$9. Sticker-shock take: the cheap tier got expensive. Ch 29's read: on the boards Google showed, that Flash cleared last-gen 3.1 Pro on agentic work — a model that one-shots what the old one needed three turns and a retry for is cheaper per task at 3× the per-token price. The sticker and the task cost moved in opposite directions, and nothing on the pricing page told you which way it broke for your workload.

Fable 5 is the same test at the top of the ladder instead of the bottom. The launch receipt — vendor-curated, so treat it as a signal, not a private eval — is Stripe running a 50-million-line Ruby codebase migration in one day against a two-month by-hand estimate. Read that as turn-count collapse, not as magic: every turn a model doesn't need is context you don't re-send, a retry you don't pay for, a wrong answer you don't spend your evening reading. A model that finishes long-horizon work in one pass at 2× the rate beats one that needs five passes at 1× — on the invoice and on your calendar.

The launch table points the same direction — SWE-Bench Pro 80.3% against Opus 4.8's 69.2% — but the book's discipline holds: Berkeley RDI reward-hacked eight major agent benchmarks in April, the standing discount on public scores is 10–15 points, and a benchmark is never the receipt. The head-to-head case is on vs Opus 4.8; your private eval is the only number that settles it.

At a $10 input sticker, prompt caching stops being an optimization and becomes the price. Cache reads run roughly 0.1× the input rate — so the stable prefix of every Fable 5 prompt costs about $1 per million tokens instead of $10. That's Haiku-input territory for the part of your prompt that never changes, on the most capable model Anthropic sells.

And there's a quiet spec change that matters more than it looks: the minimum cacheable prefix on Fable 5 is 2048 tokens, down from 4096 on Opus 4.8. A 3K-token system prompt — a lean CLAUDE.md, a small tool schema set — caches on Fable 5 and silently doesn't cache on Opus 4.8. Silently is the operative word: no error, no warning, you pay full input price on every call and find out on the weekly invoice. If you trimmed your prompts to be frugal on Opus, Fable 5 may be the first model that pays you back for it.

• Bulk classification and extraction. Haiku 4.5 is $1/$5 — a tenth of Fable's input rate. Routing a triage queue to the frontier tier was wrong at $5/$25 and it's twice as wrong now. The tier list routing holds: cheap tier for cheap work.
• Cyber- and bio-adjacent work. Fable 5's classifiers gate offensive-cyber and most bio/chem requests. The billing is honest — a blocked API request returns an error you aren't charged for, and the opt-in fallback continues on Opus 4.8 at Opus pricing ($5/$25), with fallback built in on the apps and Managed Agents. But if your domain is the trigger area, you're running a classifier round-trip to reach the model you could've called directly. Route to Opus 4.8 directly.
• Short single-turn tasks. The whole case for 2× is turn-count collapse on long-horizon work. A one-turn task has no turns to collapse — the sticker is the entire price, and the sticker is 2×.
• The driver's seat, when the advisor seat is enough. Fable 5 ships as an advisor model: faster, cheaper workers call it mid-task to check their plan and evaluate their work. If what you want is Fable's judgment rather than its keystrokes, buy the judgment only — workers at Sonnet rates, the $10/$50 model at the gate. Details on the API page.

You have until June 22 and it costs nothing on a paid plan. Pick the three workloads where Opus 4.8 makes you wait, retry, or babysit. Run them down both lanes — Opus 4.8 as control, Fable 5 as variant, same tasks, same prompts. Keep the transcripts. Then read one number: total billed per completed task. Not dollars per million tokens, not the benchmark delta, not the vibes after one impressive session. Tokens × rate ÷ tasks that finished — that's the number the 2× either survives or doesn't, per workload, and it's different for each one.

That's not a takeaway, it's a meter installation. Mine is already running.