Claude Fable 5: Anthropic's Mythos-Class Model (2026)

Claude Fable 5 is Anthropic's most capable generally available model, launched June 9, 2026 as a "Mythos-class" system made safe for public use.¹ It posts state-of-the-art scores on nearly every benchmark Anthropic tested, but it ships with a notable asterisk: on prompts about cybersecurity, biology, chemistry, or model distillation, Fable 5 silently hands the conversation to the less-capable Claude Opus 4.8 instead.¹²

That trade — frontier capability for most users, a quiet downgrade for a few sensitive domains — is the whole story of this release. Fable 5 is the public face of the same model Anthropic calls Claude Mythos 5, the tier above its Opus class, and the direct descendant of the Mythos Preview that found thousands of zero-day vulnerabilities under Project Glasswing.¹ The capability that "spooked the government" is now, with guardrails, available to everyone.

TL;DR

Claude Fable 5 is the first publicly available Mythos-class model from Anthropic — a tier above Opus — released June 9, 2026.¹ It is the same underlying model as Claude Mythos 5; the only difference is safeguards.¹ On Anthropic's launch benchmarks, Fable 5 reports 95.0% on SWE-bench Verified, 80.3% on SWE-bench Pro, and 88.0% on Terminal-Bench 2.1, ahead of Claude Opus 4.8 (88.6% and 69.2% on the two SWE-bench tests) and the rest of the field.³⁴ Pricing is $10 per million input tokens and $50 per million output tokens — double Opus 4.8 and "less than half the price of Claude Mythos Preview."¹ The catch: Fable 5's safety classifiers route prompts touching cybersecurity, biology and chemistry, or distillation to Opus 4.8 instead, which Anthropic says happens in fewer than 5% of sessions on average (so more than 95% of sessions run on the full model).¹² For biology and chemistry the fallback is deliberately broad, and on AWS those routed requests are billed at Opus rates.¹² Fable 5 is available today across the Claude API, Claude apps, Amazon Bedrock, the Claude Platform on AWS, Vertex AI, Microsoft Foundry, GitHub Copilot, and Databricks; the unrestricted Mythos 5 stays locked to vetted Glasswing and life-sciences partners.²⁵

What is Claude Fable 5?

Claude Fable 5 is a frontier large language model from Anthropic, released June 9, 2026, that the company describes as "a Mythos-class model that we've made safe for general use."¹ In Anthropic's lineup, Mythos-class sits one tier above the Opus class in raw capability. The first Mythos-class model, Claude Mythos Preview, shipped in April 2026 to a small group of cyber defenders through Project Glasswing; Fable 5 and Mythos 5 are the second generation, and the first time these capabilities reach the general public.¹

Anthropic's framing is blunt: "Fable 5's capabilities exceed those of any model we've ever made generally available… The longer and more complex the task, the larger Fable 5's lead over our other models."¹ It is built for long-running, asynchronous work — coding and knowledge tasks that run for extended periods without human intervention — with stronger vision and the ability to develop its own harnesses, evaluations, and skill updates as it goes.² The name is a tell, too: Anthropic notes that "Fable" comes from the Latin fabula, "that which is told," akin to the Greek mythos — the safeguards are what separate the two models.¹

Fable 5 vs Mythos 5: one model, two sets of safeguards

The single most important thing to understand about this launch is that Fable 5 and Mythos 5 are the same underlying model.¹ What differs is the guardrails. Mythos 5 is the unrestricted version — "the same model as Claude Fable 5 but with cyber safeguards lifted" — and it is not for sale to the public.¹ It goes only to vetted partners: Glasswing cybersecurity organizations (in collaboration with the US government), where it serves as an upgrade to Mythos Preview, and, soon, a small group of biology researchers through a separate trusted-access program.¹ Anthropic says Mythos 5 "has the strongest cybersecurity capabilities of any model in the world."¹

Fable 5 is that same engine with a classifier layer bolted on. When you query it on a sensitive topic, a separate AI classifier intercepts the request and a response comes back from Claude Opus 4.8 instead — Anthropic's next-most-capable model — and you're told when it happens.¹ For more than 95% of sessions, no fallback occurs and "Fable 5's performance is effectively the same as that of Mythos 5."¹ For the rest, you're talking to Opus 4.8 without quite realizing the frontier model stepped aside.

Claude Fable 5 benchmarks

On Anthropic's launch benchmarks, Fable 5's coding numbers are the strongest the company has published. (Anthropic released its comparison table as an image; the figures below come from that table as transcribed in launch-day coverage, and — like any vendor's launch numbers — are self-reported.)³⁴

The standout is 95.0% on SWE-bench Verified — fixing real bugs in real GitHub repositories — which would be the highest score on that test to date, above Claude Opus 4.8's 88.6% and GPT-5.5's 88.7% (OpenAI-reported).³ On SWE-bench Pro, the harder, contamination-resistant variant, Fable 5's 80.3% sits roughly 11 points clear of Opus 4.8 (69.2%) and more than 20 ahead of GPT-5.5 (58.6%) and Gemini 3.1 Pro (54.2%).⁴ It also reports 88.0% on Terminal-Bench 2.1, 59.0% on Humanity's Last Exam without tools, and tops 90% on Hex's analytical benchmark — "a 10-point jump over Opus," per Hex's own testing.¹⁴

The usual caveat applies harder than usual here, because the official table is an image rather than machine-readable data: these are Anthropic's own numbers, and SWE-bench-style scores vary by harness (GPT-5.5, for instance, lands anywhere from ~82% to its official 88.7% depending on the evaluation setup).³ Treat the headline as a strong, self-reported claim pending independent replication.

What Claude Fable 5 can do

Where the release gets more concrete is the named-customer evidence Anthropic published alongside the benchmarks. In software engineering, Anthropic says Stripe used Fable 5 to run a codebase-wide migration across a 50-million-line Ruby codebase in a single day — work it estimates would have taken a team over two months by hand.¹ Cognition reported Fable 5 as the highest-scoring model on its FrontierCode evaluation, Cursor called it state-of-the-art on CursorBench, and Replit said it topped its end-to-end ViBench vibe-coding test.¹

Beyond code, Anthropic highlights vision and long-horizon endurance. Fable 5 is pitched as a new vision state-of-the-art — it rebuilt a web app's source code from screenshots alone, and beat Pokémon FireRed using a minimal, vision-only harness where earlier Claude models needed elaborate scaffolding.¹ On long-context work, giving the model file-based memory while it played the deck-builder Slay the Spire improved its performance three times more than the same memory helped Opus 4.8.¹ One early tester, physics firm CEO Matthew Pines, said Fable 5 "in 36 hours got nearly to where GPT-5.5 landed after four days," while using a third of the reasoning tokens.¹ These are Anthropic-selected, early-access reports rather than independent tests — but they're specific, named, and consistent with the benchmark story.

The most striking claims belong to the unrestricted Mythos 5: Anthropic's internal teams used it to accelerate parts of drug design roughly tenfold, and it ran more than a week of largely autonomous genomics research, training a custom model that — at a hundredth the size — beat a recent model published in Science.¹ That is the capability ceiling Fable 5 inherits, minus the dangerous-domain access.

The catch: when Fable 5 silently becomes Opus 4.8

Here is the part that matters for anyone evaluating Fable 5 for real work. Its safeguards are not a content filter that refuses you — they're a routing layer that swaps the model. Anthropic's classifiers cover three areas: cybersecurity, biology and chemistry, and distillation (attempts to extract Fable's capabilities to train competing models).¹ Trip one, and Opus 4.8 answers instead.¹

For cybersecurity, that's a deliberate wall: Anthropic says the classifiers "prevent Fable from making any progress" on offensive cyber tasks, and one external partner found Fable 5's cyber safeguards the most robust of any model it tested, including Opus 4.7 and 4.8.¹ An external bug bounty ran more than 1,000 hours without producing a universal jailbreak, though Anthropic notes the UK AI Safety Institute "has made progress towards one."¹ For biology and chemistry, the fallback is intentionally broad — Anthropic has "arranged for Fable to fall back to Opus 4.8 on most requests" in those domains while it tunes the classifiers.¹

The implication is uncomfortable for exactly the users who'd most want a Mythos-class model. A security researcher probing their own systems, or a biologist doing legitimate protein work, will frequently find their "Fable 5" sessions quietly served by Opus 4.8 — and on Amazon Bedrock, those routed tokens are billed at Opus rates, with mid-conversation blocks charging Fable rates up front and Opus rates after.² Anthropic is candid that the safeguards are "stricter than would be ideal" and "sometimes catch harmless requests," and says it will narrow them over time.¹ There's also a privacy cost: Anthropic now requires 30-day data retention for all Mythos-class traffic (it says the data won't be used for training and will be deleted afterward), and on AWS you must opt into data sharing before you can call the model at all.¹²

Claude Fable 5 pricing and availability

Fable 5 and Mythos 5 are both priced at $10 per million input tokens and $50 per million output tokens.¹ That's exactly double Claude Opus 4.8's $5/$25, but Anthropic frames it as a bargain relative to the tier — "less than half the price of Claude Mythos Preview."¹ When a prompt falls back to Opus, you pay Opus prices for the routed portion, not Fable's.²

Access is broad but, for subscribers, time-boxed. On the Claude API and consumption-based Enterprise plans, Fable 5 is fully available now.¹ For subscription tiers, Anthropic is rolling out in stages because it expects demand to outstrip capacity: from June 9 through June 22, Fable 5 is included on Pro, Max, Team, and seat-based Enterprise plans at no extra cost; on June 23 it leaves those plans and further use requires usage credits, until Anthropic can restore it as a standard inclusion.¹ In other words, the "free" window is a two-week countdown.

The model is live across an unusually wide surface for a launch day: the Claude API (as claude-fable-5), Claude's apps, Amazon Bedrock (US East and Europe regions at launch) and the Claude Platform on AWS, Google's Vertex AI, Microsoft Foundry, GitHub Copilot, and Databricks.¹²⁵ The unrestricted Mythos 5 remains a limited preview for vetted cyber and life-sciences partners only.¹²

From Mythos Preview to Fable 5: the Glasswing throughline

Fable 5 only makes sense against the backdrop of the past two months. In April, Anthropic released Claude Mythos Preview to a tiny group of defenders under Project Glasswing, after the model demonstrated a superhuman ability to find and exploit software vulnerabilities — the work that drew government attention and seeded the blog's earlier coverage of Mythos-class cyber capabilities. At the time, Anthropic said it hoped to eventually release Mythos-level capability to all users if it could build safeguards strong enough to prevent misuse.¹

Fable 5 is Anthropic declaring those safeguards good enough — for now. It arrives the same week the open-weight world is racing in the opposite direction, with models like MiniMax M3 and Nex-N2-Pro shipping frontier-adjacent coding capability with no guardrails and downloadable weights. Fable 5 is the bet that the most capable models can be both broadly available and gated; the open-weight wave is the bet that they can't, and won't be. This launch is the clearest test yet of which model of distribution wins.

Bottom line

Claude Fable 5 is a genuine milestone — the first time Anthropic has put Mythos-class capability in front of the general public — and its self-reported numbers, led by a 95% SWE-bench Verified score, would make it the most capable coding model available if they hold up independently.¹³ But the headline isn't the benchmark; it's the safeguard. Fable 5 is a frontier model with a trapdoor: on the very domains where its capability is most valuable and most dangerous, it steps aside for Opus 4.8, and it asks for broad data retention as the price of admission.¹² For most developers building agents and doing long-horizon work, that trapdoor will never open and Fable 5 will simply be the best model they can buy. For security researchers and life scientists, the model they're paying a premium for will often be a different, weaker one. Whether that's a reasonable price for releasing this much capability safely — or an awkward halfway house — is the question Anthropic has now put to the whole market.

Related reading: Claude Mythos Preview and Project Glasswing, Claude Opus 4.8: benchmarks and pricing, and Nex-N2-Pro: the open-weight challenger.