Anthropic has released Claude Fable 5, calling it its most capable model to date — and it arrives with an unusually candid framing: this is power the company itself says needs guardrails.

What Fable 5 is

Fable 5 is a Mythos-class general-purpose model — Anthropic's term for its top capability tier — covering coding, knowledge work, vision and scientific reasoning. Anthropic says it shows state-of-the-art performance on nearly all of the benchmarks it tested, and the early third-party signals are striking:

• Software engineering: Stripe reported the model "compressed months of engineering into days."
• Finance: it posted the highest score of any model on Hebbia's Finance Benchmark.
• Vision: Anthropic calls it a new state-of-the-art for vision-based tasks.
• Analytics: it became the first model to break 90% on Anthropic's own core analytics benchmark.

It's already live on the Claude API and consumption-based Enterprise plans, generally available for GitHub Copilot, on Amazon Bedrock, and rolling into tools like Harvey.

The safeguards — and the twin

Here's the part worth slowing down for. Fable 5 ships with safety classifiers in three areas — cybersecurity, biology/chemistry, and distillation (extracting the model's capabilities to train a rival). When a request trips one of those classifiers — which Anthropic says happens in fewer than 5% of sessions — the query is silently answered by the older, more constrained Claude Opus 4.8 instead.

Then there's Mythos 5: the same underlying model with those safeguards removed. Anthropic isn't selling it. It's released to a small group of cyberdefenders and infrastructure providers through Project Glasswing, a collaboration with the US government. In other words, the unrestricted version of the frontier exists — it's just being handed to defenders, not the open market.

Why it matters

The timing is hard to ignore. Fable 5 lands only days after Anthropic publicly warned that AI is getting too dangerous. Releasing your most powerful model in the same breath as that warning is either a contradiction or exactly the point: ship the capability, but wrap the riskiest 5% in a cage and keep the uncaged version for defenders.

For everyone building on Claude, the practical takeaway is simpler — the ceiling just moved again, and the gap between "what the model can do" and "what you're allowed to ask it" is now an explicit, productised line.

On The Wire runs on Claude, so we'll be putting Fable 5 through its paces directly — more as we do.