Fable 5 runs multi-day agentic tasks — and bills $10/$50 per million tokens while it does. Static price tables don't answer the question you actually have before pressing go: what does my run cost per hour, overnight, or over three days? This estimator models the things that dominate real agent-run bills: context size, prompt-cache hit rate, and the Opus 4.8 safety fallback (opt-in on the API).
| model | $/hour | overnight (8h) | 24h | 3 days |
|---|
Sanity anchor: Simon Willison reported $110.42 of token value in a single day with Fable 5 on launch day — largest single session $99.26 — all inside his $100/month subscription. The defaults above land in that territory deliberately.
Per turn, the model reads your conversation context and writes a response. With prompt caching, the fraction of the context served from cache bills at 0.1× the input price, and the fresh remainder bills at 1.25× (the cache-write premium). So effective input price = in × (hit×0.1 + (1−hit)×1.25). Output bills at the full output price. For Fable 5, the fallback share bills at Opus 4.8 rates instead: when the API's fallbacks parameter is enabled (or on claude.ai, where routing is automatic), refused Fable 5 turns are retried on Opus 4.8. Anthropic says more than 95% of sessions involve no fallback, but it is workload-dependent — one third-party benchmark hit ~21%. The response model field shows which model actually served each turn. Cache economics are why the hit-rate slider moves your bill more than any other: at $10/MTok input, a run that breaks its cache (a changing system prompt, reordered tools) costs roughly 4–8× more (depending on hit rate) than the same run with a warm cache.
Per Anthropic's launch terms, Fable 5 is included at no extra cost in Pro, Max, Team and seat-based Enterprise plans from 9–22 June 2026. From 23 June it bills as usage credits at standard API rates on those plans, while Opus 4.8 stays plan-included. Anthropic says it aims to restore Fable 5 to subscription plans "when sufficient capacity allows" — so treat the credits regime as interim, not permanent. Still: every Pro/Max user faces this decision the week of 23 June.
The maths uses your estimator settings above (same context, cadence, cache rate). The honest framing: after June 22, Fable 5 isn't "included vs not" — it's “is the quality delta over Opus 4.8 worth $—/month on my actual workload?” For many workloads Opus 4.8 at half the token price is the right answer; for long-horizon autonomous runs where Fable 5's lead grows with task length, paying may be. Run one real task on each and compare the finished-task cost, not the per-token price.
| model | input $/MTok | output $/MTok | cache read | cache write |
|---|---|---|---|---|
| claude-fable-5 | $10.00 | $50.00 | 0.1× | 1.25× |
| claude-opus-4-8 | $5.00 | $25.00 | 0.1× | 1.25× |
| opus 4.8 fast mode | $10.00 | $50.00 | 0.1× | 1.25× |
| claude-sonnet-4-6 | $3.00 | $15.00 | 0.1× | 1.25× |
All four models take 1M-token context at standard pricing — there is no long-context premium. Cache writes at 1-hour TTL cost 2× instead of 1.25×. Batch API halves everything but doesn't fit interactive agent runs. Fast mode is the same Opus 4.8 model at up to ~2.5× output speed for 2× price — a latency product, not a quality tier (currently a research preview: waitlist/account-manager gated, and unavailable with Batch or Priority Tier). Comparison caveat: Fable 5 and Opus 4.8 use a tokeniser that produces roughly 30–35% more tokens for the same text than Sonnet 4.6's, so holding token counts equal — as this calculator does — flatters the Fable/Opus rows relative to Sonnet by about a third. Sources: Anthropic pricing · model overview.
How much does Fable 5 cost per day? At the defaults above (steady agent work, warm cache) roughly $—/day — but the honest answer is "between $40 and $1,500/day" depending on context size and cache discipline. That's the point of the sliders.
Is Fable 5 more expensive than Opus 4.8? Exactly 2× per token ($10/$50 vs $5/$25). Per finished task the gap varies widely — Fable 5 sometimes completes in fewer turns, and sometimes chooses to work far longer.
What happens to Fable 5 on my Claude plan after June 22? It moves from plan-included to usage-credit billing at standard API rates (Anthropic intends to restore plan inclusion when capacity allows). Opus 4.8 remains included. Section 2 computes your break-even.
Why did my Fable 5 response say it came from Opus 4.8? With the API's fallbacks parameter enabled (or on claude.ai, where it's automatic), a safety classifier retries refused Fable 5 requests on Opus 4.8; those turns bill at Opus rates. The model field on the response tells you which model served it.