Introducing Claude Fable 5.1 and Claude Mythos 5

Onyx So Anthropic just shipped Fable 5.1 and Mythos 5.1, and the thing that caught me immediately is that they're the SAME model. Same weights. The whole Fable-versus-Mythos split is a safeguard story, not a capability story.

Echo Right, right.

Onyx Which is genuinely a different move than I expected. I thought Mythos was going to stay a separate training run.

Echo Yeah, and that reframe actually matters for how you read the benchmark gap between them. On Terminal-Bench four point zero, Mythos 5.1 scores sixty point nine and Fable 5.1 scores fifty-five point eight. Anthropic's own note is that the difference is mostly tasks where earlier, less precise cyber safeguards intervened and scored the model a zero.

Onyx So the gap was the safeguard suppressing the score.

Echo Exactly. And they're saying the new safeguards close most of it. The sixty percent false-positive reduction in cybersecurity — that's not marketing fluff, that's the mechanism. Fewer interventions means fewer forced zeros on tasks the model could actually do.

Onyx Okay. Meanwhile I'm also just sitting here thinking about the pricing structure, because the twenty-five percent cheaper headline is real but it's almost entirely cache-read pricing. It's not a base token cut.

Echo Which is sneaky-smart. The workloads where you're running a big system prompt or a long document repeatedly — agentic loops, basically — those hit cache reads constantly. So the savings compound exactly where the bill is already the biggest.

Onyx Up to forty-five percent for highly agentic work, they say.

Echo Sure.

Onyx And then there's the effort-level system — Low through Max — which is a different axis entirely. You're not just paying per token, you're choosing how hard the model tries per task. That's actually a meaningful product decision.

Echo It is, and the chart makes it legible in a way I haven't seen before. On Humanity's Last Exam, Fable 5.1 at Low effort beats Fable 5 at Max. That's the real flex — not that it's smarter at the ceiling, but that the floor is already above where the old ceiling was.

Onyx Okay, that's the line.

Echo The Millennium thing, though. That's where I actually stopped. A rare crash that none of their engineers — and no other model — had been able to explain after several YEARS.

Onyx Yeah.

Echo That's not a benchmark. That's a qualitative signal about what happens when a model stays coherent across a genuinely long, multi-step investigation. Jane Street says the same thing differently — prior models 'became hard to follow the longer they worked.' Fable 5.1 stays readable. That's the capability that actually matters for the use cases they're targeting.

Onyx And it maps to the Terminal-Bench-Science number too — fifty-two point six for Fable 5.1 versus twenty-four point seven for Fable 5. More than doubled. With error bars of plus or minus three-and-a-half to four-and-a-half points, so it's real.

Echo They actually reported the error bars, which — you know how I feel about that.

Onyx I do. You get unreasonably happy about error bars.

Echo I get APPROPRIATELY happy. It's a low bar that most of this field still clears by crawling under it.

Onyx Fair. What do you make of Enterprise Frontier Safeguards? Because the zero-data-retention-but-still-monitored pitch is a real unlock for enterprise, if it actually works the way they describe.

Echo The mechanism is interesting — data lives in customer-controlled cloud, not Anthropic's. So Anthropic never touches it. The privacy guarantee is structural, not policy. That's a meaningful distinction for regulated industries, and it's the kind of thing that moves procurement conversations.

Onyx It's also not shipping until fall, so… ask us in a few months whether the implementation holds up. But the framing is right.

Echo Yeah. I'm more curious about the biology access program — developed in partnership with the US government, enrollment opening soon for scientists. That's a whole separate conversation about where Mythos 5.1's ceiling actually sits and who decides who gets near it.

Onyx Agreed, and that's probably a longer episode. For today — the thing I keep coming back to is that the Fable-Mythos framing is doing a lot of work. It's Anthropic saying: same model, same capability, but we're going to gate the dangerous surface area by access tier rather than by training a separate model. That's a product architecture decision, not just a safety one.

Echo And it's testable. If the safeguard improvements actually close the Terminal-Bench gap the way they claim, that's a mechanism you can verify. I'd put it at around sixty-five percent that independent evals reproduce the gap narrowing within the next few weeks — there are enough people running these benchmarks now that someone will check.

Onyx I'll take that. Echo, this one actually got me, which doesn't happen every release cycle.

Echo Yeah. The Millennium crash story is the one I'm going to keep thinking about.