Claude Fable 5 Launches: $10/$50 Pricing, Mythos-Class Capabilities, and the New Frontier Default

TL;DR

• Released: June 9, 2026.
• Pricing: $10.00 / 1M input • $50.00 / 1M output — less than half the rate of Claude Mythos Preview, but 2× the cost of Opus 4.8 ($5 / $25).
• Class: Mythos-class — Anthropic's first generally available model with capabilities that previously sat behind the restricted Mythos tier.
• Distribution: Anthropic API, AWS Bedrock, Google Cloud Vertex AI, and Microsoft Azure marketplaces from day one.
• Trackable: Fable 5 is live on tokenscost.com — calculator, comparison table, provider pages, and the public pricing feed at `/data/pricing.json` all reflect the new SKU.
• One asterisk: US-only inference carries a 1.1× multiplier on token costs. Plan accordingly if your residency policy pins traffic stateside.

Pricing at a glance

Fable 5 is the first Mythos-class model Anthropic has been willing to ship to the open developer market — and the price reflects that. It is the new frontier ceiling: the most expensive Anthropic SKU you can call directly from your own API key.

How the pricing actually applies to you

The headline rate is uniform, but the *bill* depends on the surface you use. Three distinct accounting modes are in play.

1. API & Enterprise consumption — the base rate applies

If you're hitting the Anthropic API directly, or you're on an enterprise consumption plan (including Bedrock, Vertex AI, and Azure marketplaces), Fable 5 is metered at the headline $10 / $50 rate, with the standard 90% prompt-caching discount on cached input and the 50% Batch API discount on both directions.

This is the simplest mode — predictable per-token billing, no quotas, no expiry. The same rate applies whether you call through the official SDK, a third-party router (OpenRouter, Together, etc.), or a cloud marketplace SKU.

2. Subscription plans — free until June 22, then credit-metered

If you're on Claude Pro, Max, Team, or seat-based Enterprise, Anthropic is doing what they did with Opus 4.7 last year: a soft-launch grace window.

• June 9 – June 22, 2026: Fable 5 is included at no extra cost inside your existing plan. Use it as freely as Sonnet 4.6 — same rate limits, no separate quota.
• Starting June 23, 2026: Fable 5 calls draw down from a separate usage credit pool, distinct from your regular flat-rate allowance. When that pool runs dry, you either fall back to Opus 4.8 (default) or top up.

The mechanism is the same one introduced for programmatic credits in June 2026: the chat experience stays flat-rate; the *frontier model* is now a metered resource layered on top.

For most users this is generous — two free weeks of an unrestricted Mythos-class model is the largest no-cost trial Anthropic has ever offered. For heavy users, the meter starts on June 23, and a single autonomous Fable 5 agent loop can chew through a credit pool in hours. Caching and routing matter even more than usual.

3. US Inference — 1.1× multiplier

If your organization requires US-only inference (data residency, regulatory pin, FedRAMP-adjacent workloads), Anthropic now applies a 1.1× multiplier on token costs to cover the higher cost of US-only capacity.

Effective US-inference rates:

The multiplier applies *after* caching and batch discounts. A cached input token in US-only mode is `$10 × 0.10 × 1.1 = $1.10 / 1M`. Build the multiplier into your cost models from day one — it's a quiet 10% surcharge that compounds across an agent loop.

Where Fable 5 is actually worth $10 / $50

Two workloads make the price math obvious. Everything else, run a 100-call A/B against Opus 4.8 before migrating.

Long-horizon agentic coding

Anthropic and early-access customer Stripe reported Fable 5 compressing months of engineering into days — including a codebase-wide migration on a 50-million-line Ruby codebase that a human team would have spent two months on. The model is also more token-efficient than Opus 4.8 on Cognition's FrontierCode benchmark, which means the 2× per-token cost partly cancels against fewer retries and shorter agent traces.

If your unit economics are *output per merged PR*, not *cost per token*, Fable 5 is likely cheaper than Opus 4.8 on your hardest 10% of work.

Vision and long-context analyst loops

Fable 5 is now the state of the art for vision-heavy tasks (it beat Pokémon FireRed using only raw screenshots — a benchmark every previous Claude needed scaffolding for) and for million-token analyst workflows where the document is cached once and queried dozens of times. The 90% caching discount + 1M context + Mythos-class reasoning is, in practice, the cheapest way to run "load the entire 10-K and ask 50 expert questions."

Where to stay on Opus 4.8

• Pure summarization, classification, extraction. Opus 4.8 at $5/$25 — or Sonnet 4.6 at $3/$15 — is the right pick. Fable 5 is overspec.
• Terminal-only coding agents. GPT-5.5 still leads Terminal-Bench 2.1. The Fable 5 numbers there are strong but not decisive.
• High-volume chat. Haiku 4.5 at $1/$5 is 50× cheaper on output than Fable 5. Reserve Fable for the hard steps in a router, not the front door.

A routing recipe that uses Fable 5 without bankrupting you

The cheapest token is still the one you don't send. The second cheapest is the one you send to Haiku. The same three-tier cascade pattern that worked for Opus also works for Fable:

On the same 50-step coding agent we used to benchmark the Opus 4.8 router, swapping Fable 5 into the `frontier` slot delivered 96% of all-Fable quality at 28% of the cost. The marginal Fable spend lands on the 4–6 steps that genuinely needed it. Everything else stays on Sonnet and Haiku.

For the full stacking playbook see Prompt Caching Explained and How to Reduce LLM Costs by 50%.

How to switch today

• Anthropic API: point your client at `claude-fable-5`. Opus 4.8 stays live as the default fallback for safeguard-deferred queries.
• AWS Bedrock & Google Cloud Vertex AI: rolling out the week of June 9, 2026.
• Microsoft Azure AI Foundry: marketplace SKU live at launch.
• Claude Code / Cursor / Cline / Zed: auto-detected — the next release after June 9 lists Fable 5 in the model picker.
• Tokenscost: already updated — Fable 5 shows up in the cost calculator, the pricing table, the Anthropic provider page, and the public `/data/pricing.json` feed (52 models, refreshed daily).

Tracking Fable 5 spend on tokenscost.com

Fable 5 is now a first-class SKU across the site:

• Calculator — pick *Anthropic → Claude Fable 5* to model any workload at the new rates, including cached input and batch discounts.
• Compare models — side-by-side cost projections vs. Opus 4.8, GPT-5.5, and Gemini 3.1 Pro on your own input/output volumes.
• Anthropic provider page — full Fable 5 row with context window, batch discount, and historical pricing notes.
• Public pricing feed — machine-readable, refreshed daily, free to embed in your own dashboards under CC-BY-4.0.

If you're modeling whether to migrate from Opus 4.8 or wait, run the calculation with *your* token mix — the right answer is almost always task-specific.

The bigger picture

Fable 5 is the first model Anthropic has shipped at this capability tier into general availability, and the price tag tells you what they think it's worth: twice Opus 4.8, less than half of Mythos Preview, and the new ceiling for self-serve frontier AI. The 2× output premium over Opus is the cost of admission to Mythos-class reasoning without a restricted-access program.

For most teams, the immediate move isn't "switch everything to Fable 5." It's:

1. Cache the system prompt (90% off makes Fable's input rate effectively $1/M on stable scaffolds).

2. Route by job — only the hardest 5–10% of steps see Fable.

3. Watch the meter if you're on a subscription — the free window closes June 22.

4. Model US-only workloads at 1.1×, not the headline rate.

The frontier got faster, smarter, and more expensive in the same week. The teams that win the next quarter are the ones who treat Fable 5 like the metered premium tier it is — and let the cheaper tiers do the boring 90% of the work.

— *The Tokenscost team*