Claude Fable 5: Anthropic Releases Its Most Powerful Model to the Public

TL;DR: Anthropic just dropped Claude Fable 5 – a public version of their restricted Mythos-class model that was previously locked behind Project Glasswing. It tops nearly every benchmark, costs $10/$50 per million tokens (input/output), and is available to everyone starting today. The catch: ~5% of sessions trigger safety fallbacks to Opus 4.8, and subscription users only get free access until June 22.

What Just Happened

On June 9, 2026, Anthropic released two models simultaneously: Claude Fable 5 (for everyone) and Claude Mythos 5 (for approved cyber defenders and researchers only). Both are built on the same underlying “Mythos-class” architecture – the same model family that Anthropic said was too dangerous to release publicly just two months ago.

Now it’s public. With guardrails.

Here’s what changed: Anthropic developed a new layer of safety classifiers that automatically route risky queries (cybersecurity, biology/chemistry, model distillation) to Claude Opus 4.8 instead. Over 95% of sessions see zero fallbacks. The other 5% get Opus 4.8 responses – still a strong model, just not the bleeding-edge one.

The Numbers That Matter

Benchmark	Fable 5 / Mythos 5	Opus 4.8	GPT-5.5
SWE-bench Pro (coding)	80.3%	–	58.6%
FrontierCode Diamond	29.3%	13.4%	5.7%
Finance Benchmark (Hebbia)	#1	–	–
Legal Agent Benchmark	#1	–	–
CursorBench (coding)	#1	–	–

Those FrontierCode numbers are wild. Fable 5 at 29.3% vs GPT-5.5 at 5.7% – that’s a 5x gap on production-quality coding. And on SWE-bench Pro, Fable 5 beats GPT-5.5 by over 20 points.

The real-world test that stands out: Stripe ran Fable 5 on their 50-million-line Ruby codebase. A migration that would have taken an engineering team two months was completed in a single day. That’s not a benchmark – that’s a fortune 500 company putting real money behind the claim.

Pricing – The Elephant in the Room

Model	Input (per 1M tokens)	Output (per 1M tokens)	Total Cost
DeepSeek V4 Flash	$0.14	$0.28	$0.42
Gemini 3.5 Flash	$1.50	$9.00	$10.50
GPT-5.4	$2.50	$15.00	$17.50
Claude Opus 4.8	$5.00	$25.00	$30.00
GPT-5.5	$5.00	$30.00	$35.00
Claude Fable 5	$10.00	$50.00	$60.00

Fable 5 is the most expensive generally available AI model on the market. At $60 per million tokens total, it costs nearly double GPT-5.5 and twice Opus 4.8. But it’s less than half the price of the old Mythos Preview – so Anthropic is framing this as a price cut, not a premium.

For subscription users (Pro, Max, Team, Enterprise): Fable 5 is included free through June 22. After that, you’ll need usage credits. Anthropic says they’re working to bring it back to subscription plans “as quickly as possible.” Translation: expect to pay extra.

What Fable 5 Can Actually Do

Coding: Autonomous multi-hour coding sessions. Codebase-wide migrations. Production-quality code that passes real test suites. Stripe’s 50M-line migration is the headline, but Cursor, Cognition (Devin), and GitHub all confirmed it’s the best coding model they’ve tested.

Vision: State-of-the-art. Can rebuild web apps from screenshots. Beat Pokemon FireRed using only raw game screenshots – no maps, no navigation aids, no game state data. Previous Claude models couldn’t do this even with helper tools.

Finance: Highest score ever on Hebbia’s Finance Benchmark. IMC (the trading firm) said it “aced their trading-analysis evaluations nearly across the board.” If you’re using AI for financial analysis, this is the best model available.

Science: Mythos 5 (the unrestricted version) accelerated drug design workflows by 10x. It autonomously designed protein therapeutics that matched or beat skilled human scientists. One of its molecular biology hypotheses was independently confirmed by a separate research lab.

Long-context: 1M token context window. Stays focused across millions of tokens in long-running tasks. In Slay the Spire testing, giving Fable 5 persistent memory improved its performance 3x more than Opus 4.8 – meaning it actually learns from its own notes better than previous models.

The Safety Architecture – And Why It Matters

This is where Fable 5 gets interesting from a business perspective. Anthropic is pioneering a new release pattern: instead of refusing dangerous queries outright, they route them to a less capable model (Opus 4.8). You still get an answer – just not the nuclear-grade one.

Three areas trigger the safety classifiers:

1. Cybersecurity – offensive hacking, exploit development, defense evasion
2. Biology and chemistry – anything that could relate to bioweapons design
3. Model distillation – attempts to extract Fable 5’s capabilities into other models

Anthropic says 1,000+ hours of external red-teaming found zero universal jailbreaks. The UK AI Safety Institute made “progress towards one” but didn’t crack it in their testing window. That’s a bold claim – and it’ll be tested aggressively by every security researcher and bad actor on the internet starting today.

Who Should Care

Developers and AI engineers: If you’re building agents, Fable 5 is the new ceiling. The autonomous coding capabilities are genuinely a generation ahead. The API ID is claude-fable-5.

Enterprise buyers: The Stripe case study alone should get budget approved. But watch the pricing – at $60/M tokens, costs will add up fast for high-volume workflows.

Investors: Anthropic is now clearly ahead of OpenAI on coding benchmarks. The 5x gap on FrontierCode Diamond is not close. But they’re also charging a massive premium – the question is whether enterprises will pay double for a model that’s measurably better, or stick with cheaper alternatives that are “good enough.”

Everyone else: If you have a Claude Pro or Max subscription, try it before June 22 while it’s free. After that, you’ll need credits.

BetOnAI Verdict

Fable 5 is the real deal – and honestly, it’s exciting to see the AI race hit this level. The benchmark numbers are best-in-class across the board, the Stripe deployment proves it works at massive scale, and the safety architecture is a smart approach to releasing powerful capabilities without being reckless about it.

A few things worth noting:

1. Price will come down. $60/M tokens is premium today, but that’s the pattern with every frontier model – expensive at launch, cheaper within months. Anthropic already cut it to less than half of Mythos Preview pricing. Expect further drops as Sonnet and Haiku-class models inherit these capabilities.
2. The subscription situation is temporary. Free until June 22, then credits required – but Anthropic has said they’re working to bring it back to subscription plans. This is growing pains, not a permanent paywall.
3. The 5% fallback rate is actually clever. Instead of refusing requests outright, you get Opus 4.8 – still one of the best models in the world. That’s a much better user experience than a blank wall, and the rate will shrink as Anthropic refines its classifiers.

What makes this moment genuinely exciting is the competition it sets up. Anthropic just raised the bar significantly – SWE-bench Pro at 80.3% and a 5x lead on FrontierCode Diamond are numbers that will force everyone else to respond. OpenAI still has GPT-5.5 as their flagship, but with gaps this wide on coding and agentic tasks, we’re watching closely for their next move. Google’s Gemini 3.1 Pro is still in preview. The pressure is on across the board – and that’s great news for everyone building with AI.

Bottom line: If you write code, analyze documents, or build AI agents for a living – Fable 5 is the best tool available today. The AI arms race just got another gear, and we’re all better off for it.

Sources: Anthropic announcement, VentureBeat, Bitcoin.com