benchrPricing breakdown
| Tier | Rate / 1M tokens |
|---|---|
| Standard input | $10.00 |
| Standard output | $50.00 |
| Cached input | $1.00 |
| Context window | 1,000,000 tokens |
| Max output | 128,000 tokens |
| Claude plans | Free June 9–22 · credits after |
Availability is wide from day one: the Claude API, Claude Platform on AWS, Amazon Bedrock, Vertex AI, and Microsoft Foundry, all under the id claude-fable-5. Claude Mythos 5, the same model with safeguards lifted, lists at the same $10/$50 but is restricted to approved Project Glasswing organizations, so for budgeting purposes Fable 5 is the Mythos-class price.
The tokenizer surcharge nobody puts on the sticker
Fable 5 uses the tokenizer that arrived with Opus 4.7, and Anthropic's own docs note it produces roughly 30% more tokens than pre-4.7 models for the same text. That matters for one specific comparison: if your budget baseline comes from an older Claude bill (Sonnet 4.5, Opus 4.5, anything before April 2026), the same workload on Fable 5 pays twice the rate and tokenizes longer. Against Opus 4.8 or 4.7 the comparison is clean, since they share the tokenizer. Against anything older, model your costs on a 1.3× token multiplier before you commit.
Caching math at the $10 tier
The 90% caching discount matters more here than on any cheaper Claude, because the absolute savings scale with the rate. An agent with a 100K-token system prompt called 1,000 times a day: at the standard rate that prompt alone bills $1,000/day. With caching, the repeated portion bills at Fable 5's $1/1M cached rate: about $100/day plus the first-write cost. The same discount on Haiku saves you $90/day; here it saves $900. At the Mythos-class tier, cache architecture isn't an optimization, it's the difference between a viable budget and an unviable one. Model your exact mix in the cost calculator.
Fallback economics: the classifier question
Fable 5 carries safety classifiers in three areas (offensive cyber, most biology and chemistry, and distillation) and routes those sessions to Opus 4.8. Anthropic says more than 95% of sessions never trigger a fallback. What its pricing page doesn't publish is a separate billing rule for the sessions that do. Until that's documented, the conservative budget assumes the Fable 5 rate on every call. The practical advice is simpler: if a meaningful share of your workload sits near those categories, you're paying a 2× premium for answers that come from Opus 4.8 anyway — buy Opus 4.8 at $5/$25 directly instead.
Is double the price worth it?
Anthropic's launch table puts the Mythos-class pair at 80.3% on SWE-bench Pro against Opus 4.8's 69.2%, an eleven-point jump on the benchmark that best predicts long agentic coding. A coding agent burning 2M input / 400K output daily costs about $40/day on Fable 5 against $20 on Opus 4.8. If the stronger model saves one engineer-hour a week of unsticking stuck loops, the premium pays for itself several times over. If your traffic is chat, drafting, and routine summarization, it never will. That work belongs on Sonnet 4.6 or below. The launch coverage has the fuller capability picture.
The free window is the whole strategy
Fable 5 is included free on Pro, Max, Team, and seat-based Enterprise plans from June 9 through June 22, 2026, then moves to usage credits. That's a two-week, zero-cost benchmark window on Anthropic's strongest public model. Use it the way the date demands: run your hardest real workload through it this week, measure against your current model, and decide with data before the meter starts on June 23.
Use-case fit
Best for: Long-horizon agentic coding and large migrations where the SWE-bench Pro jump shows up; research-grade analysis; vision-heavy agent loops; teams whose engineer-hours cost more than the token premium.
Skip if: Your workload is routine — Sonnet 4.6 and Opus 4.8 are the rational defaults; your tasks touch security research or bio, where the classifiers hand sessions to Opus 4.8 anyway; or 30-day retention fails your compliance review.