Until today, Anthropic’s Mythos-class models were restricted to a small group of cyberdefenders and critical infrastructure providers. Claude Fable 5 changes that — it’s the same underlying model, now available to anyone, at $10/M input and $50/M output tokens.
That’s less than half the price of Mythos Preview. And the capabilities are a genuine step change.
“Stripe reported that Fable 5 compressed months of engineering into days. In a 50-million-line Ruby codebase, the model performed a codebase-wide migration in a day that would otherwise have taken a whole team over two months.”
What actually changed
New model tier, general access. Fable 5 is the first Mythos-class model available outside Project Glasswing. Mythos 5 (the same model, safeguards lifted in some areas) stays restricted to vetted cyberdefenders via Glasswing’s trusted access program.
Fallback classifiers, not hard refusals. This is the interesting part. When Fable 5’s classifiers detect a risky query — offensive cybersecurity, bioweapons-adjacent biology, distillation attempts — it doesn’t refuse. It hands the request to Claude Opus 4.8 instead. Users are told when this happens. More than 95% of sessions see zero fallbacks.
Serious capability jumps across the board:
- Software engineering: long-horizon agentic coding at a level early testers called “months into days”
- Vision: beat Pokémon FireRed using raw screenshots only — no maps, no navigation scaffolding
- Long context: stays coherent across millions of tokens; uses its own persistent notes to improve outputs mid-task
- Science: protein design 10x faster, novel molecular biology hypotheses independently corroborated in a biorXiv preprint
Pricing: $10/M input, $50/M output — less than half of Mythos Preview, cheap enough to run on real workloads.
The real story: how Anthropic shipped “too dangerous to release”
The Mythos-class threshold is where Anthropic previously drew the line. Models at this capability level could meaningfully assist cyberattackers or accelerate dual-use biology research. General release wasn’t on the table without stronger safeguards.
The classifiers are the unlock. They don’t block the model; they reroute it. Anthropic is being unusually candid: the cutoffs are deliberately conservative, benign requests will occasionally fall back to Opus 4.8, and they know that’s frustrating. The goal is to narrow false positives over time as they learn the real-world distribution.
It’s an honest framing — a lot more honest than “we take safety seriously” in a press release.
What to do
- On the API today? Fable 5 is live. Long-horizon agentic tasks are where the gap over Opus 4.8 is widest — start there.
- Using Claude Code? Early testers saw the biggest gains on complex multi-file refactors and tasks that previously needed many-turn back-and-forth. Worth trying on your hardest open tickets.
- Working in cybersecurity? Expect fallbacks to Opus 4.8. Apply for Mythos 5 access through the Glasswing trusted access program if you need the full model.
- Life sciences / biomed? The biology classifiers are intentionally broad right now. Expect friction, and track what triggers fallbacks as Anthropic narrows them.
Full capabilities and safety evaluations: Anthropic’s Fable 5 / Mythos 5 system card.
✏️ Drafted with KewBot (AI), edited and approved by Drew.
답글 남기기