Claude Fable 5 Is Here: Anthropic's Most Powerful Public Model — Benchmarks, Pricing, vs GPT-5.5
Claude Fable 5: Mythos-class capability now public. 80.3% SWE-Bench Pro beats GPT-5.5 by 21+ points with half the hallucinations. Benchmarks, $10/$50 pricing, free until June 22.
Claude Fable 5 Is Here: Anthropic's Most Powerful Public Model — Benchmarks, Pricing, vs GPT-5.5 - AI 资讯 - 一站式VPN开发平台On June 9, 2026, Anthropic officially released Claude Fable 5 — the most capable AI model the company has ever made generally available. It shares its DNA with Claude Mythos 5, released the same day under restrictions: one Mythos-class base model, two distributions. Fable 5 is the version "made safe for general use", while Mythos 5 is offered only to the U.S. government and vetted organizations through Project Glasswing.
What is Fable 5, and how does it relate to Mythos 5?
In short: one model, two release tracks.
Claude Fable 5 (public): Mythos-class capability with built-in safeguards for high-risk areas — offensive cybersecurity, biology/chemistry, and model distillation. When triggered, queries are answered by Claude Opus 4.8 instead. Anthropic says this fallback occurs in less than 5% of sessions; over 95% are handled entirely by Fable 5 itself.
Claude Mythos 5 (restricted): safeguards lifted in some areas, described as having "the strongest cybersecurity capabilities of any model in the world". Currently limited to vetted critical-infrastructure organizations and the U.S. government via Project Glasswing.
The timing is the spicy part: just days earlier, Anthropic publicly urged major AI labs to adopt "a coordinated brake pedal on frontier AI development", warning that models are approaching recursive self-improvement. Calling for brakes while shipping its most powerful public model ever made Fable 5 the most debated launch of the week.
Benchmarks: vs GPT-5.5 and Gemini 3.1 Pro
Official and third-party numbers (Vellum, Weights & Biases, and others) show Fable 5 is state-of-the-art on nearly every capability benchmark:
Software engineering (the headline)
Benchmark | Fable 5 | Opus 4.8 | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|
SWE-Bench Pro (agentic coding) | 80.3% | 69.2% | 58.6% | 54.2% |
FrontierCode Diamond | 29.3% | 13.4% | 5.7% | — |
That is a 21+ point lead over GPT-5.5 on SWE-Bench Pro, and the top score on Cognition's FrontierCode even at medium reasoning effort. Early customer Stripe reported Fable 5 "compressed months of engineering into days".
Knowledge, reasoning and vision
Benchmark | Fable 5 / Claude | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|---|
Knowledge & reasoning composite | 1932 | 1769 | 1314 |
GDP.pdf (vision, no tools) | 29.8% | 24.9% | 16.7% |
AA-Omniscience hallucination rate (lower is better) | 36.18% | 85.53% | 49.87% |
The hallucination number is the most underrated result: the Claude family hallucinates at less than half the rate of GPT-5.5 — a big deal for finance, legal and healthcare workloads. Fable 5 also posted the highest score of any model on Hebbia's Finance Benchmark.
Pricing and availability
API pricing: $10 per million input tokens / $50 per million output tokens (about half of Mythos Preview, but roughly double Opus 4.8 at $5/$25 and GPT-5.5 at $5/$30)
Max output: up to 128k tokens per request
Subscriptions: Pro / Max / Team / seat-based Enterprise get Fable 5 at no extra cost until June 22; from June 23 access moves to a credits-based model
API & enterprise platforms: generally available now (model ID:
claude-fable-5)
Safety: 1,000+ red-team hours, no universal jailbreak
High-risk topics (cyber exploitation, dangerous bio/chem research, model distillation) are blocked with automatic fallback to Opus 4.8
External red-teaming logged 1,000+ hours with no universal jailbreak found (UK AISI "made progress towards one" in an early testing window)
All Mythos-class traffic is retained for 30 days for safety purposes only — never for training
The trade-off to know: roughly one in twenty sessions is actually answered by Opus 4.8. Invisible for most workloads, but as security researchers put it — "the model you think you're using may not be the one answering".
What it means for developers and businesses
A generational gap in coding agents. 80.3% on SWE-Bench Pro means many real engineering tasks can now be delegated end-to-end. The "AI pair programmer" is becoming an "AI engineer".
Expensive per token, not necessarily per outcome. Twice GPT-5.5's price — but if it gets things right the first time, total cost can be lower, especially for large refactors with 128k output.
June 22 is a real deadline. Subscribers can use Mythos-class capability for free right now; teams evaluating it should finish their PoC within the window.
Neuronicx tracks the entire Claude product line — Claude Pro memberships, API access and other digital products are available in our marketplace, and we'll list new Fable 5 related products as they ship.