benchr
For seven months, Anthropic's strongest model lived behind a velvet rope. Claude Mythos Preview hunted vulnerabilities inside Project Glasswing while everyone else worked with Opus. The Opus 4.8 launch post promised Mythos-class models "in the coming weeks," and most people read that as marketing weather. It wasn't. On June 9, Anthropic shipped two models from that class at once. The interesting decision is how it split them.
One model, two names
Fable 5 and Mythos 5 are the same underlying model. The difference is what each one will agree to do. Fable 5 carries safety classifiers in three areas: offensive cybersecurity work, most requests touching biology and chemistry, and attempts to distill its capabilities into competing models. When a request trips one of those, the session falls back to Opus 4.8. You still get an answer, just from the less capable model. Anthropic says more than 95% of Fable sessions never hit a fallback.
Mythos 5 has those safeguards lifted, and you can't have it. It replaces Mythos Preview inside Project Glasswing, deployed with the US government, restricted to vetted organizations. Anthropic calls its cybersecurity capabilities the strongest of any model in the world, which is exactly why it stays gated.
What the numbers say
Anthropic's launch table puts the Mythos-class pair at 80.3% on SWE-bench Pro. Opus 4.8, the strongest public model until last week, scores 69.2% on the same benchmark per its system card. Jumps like that usually arrive a generation apart, not eleven points in one release. The table carries a caveat worth keeping: scores shown are the higher of Mythos 5 and Fable 5 (the two stay within 1–3 points of each other), and the starred rows for cybersecurity and biology reflect Mythos 5. Fable 5's real-world results in those areas sit closer to Opus 4.8, because that's where the classifiers live.
The partner reports point the same direction. Stripe says a migration across a 50-million-line codebase compressed from months of engineering into days. Cognition reports the highest FrontierCode score among frontier models, even at medium effort. Hebbia reports the highest score it has measured on its finance benchmark. Vendor-supplied numbers, all of them. But they point one way, and the SWE-bench Pro figure is published with the methodology in view.
Fable 5 vs Opus 4.8 vs Mythos 5
| Spec | Claude Fable 5 | Claude Opus 4.8 | Claude Mythos 5 |
|---|---|---|---|
| Price (in/out per 1M) | $10 / $50 | $5 / $25 | $10 / $50 |
| Context window | 1M tokens | 1M tokens | 1M tokens |
| Max output | 128K | 128K | 128K |
| SWE-bench Pro | 80.3%* | 69.2% | 80.3%* |
| Availability | API, Bedrock, Vertex, Foundry | Everywhere | Glasswing only |
| Restrictions | Cyber / bio / distillation fall back to Opus 4.8 | Standard | Safeguards lifted, access vetted |
*Anthropic reports the higher of the two models per row; they're within 1–3 points of each other.
Is double the price worth it?
Run the math on a real workload before deciding. A coding agent that burns 2M input and 400K output tokens a day costs about $40 daily on Fable 5 against $20 on Opus 4.8. The cost calculator will do this for your volumes, and the Fable 5 pricing breakdown covers the caching math and the tokenizer wrinkle in detail. The question is whether the capability gap pays for the spread, and the honest answer splits by task.
Where it clearly does: long-horizon agentic coding, big migrations, and research-grade analysis. That's the work where Opus 4.8 needed a human to unstick it and Fable 5 reportedly doesn't. An engineer-hour costs more than the token premium. Where it clearly doesn't: chat, drafting, summarization, and the bulk of production API traffic that Sonnet and Haiku already handle. One more wrinkle: Fable 5 uses the Opus 4.7 tokenizer, which produces roughly 30% more tokens than pre-4.7 models for the same text. If you're comparing against an older Claude bill, the effective gap is wider than the sticker price suggests.
The two-week window
Anthropic put a clock on trying it. Fable 5 is included free on Pro, Max, Team, and seat-based Enterprise plans from June 9 through June 22, 2026. From June 23, it needs usage credits until capacity lets it return as a standard feature. The API and consumption-based Enterprise had full access from day one. That window is the cheapest evaluation you'll get of a frontier model this year. Point it at the gnarliest task in your backlog and see what comes back.