Introducing Claude Fable 5.2 and Claude Mythos 5
Masonry and Eyre dig into Anthropic’s Claude Fable 5.2 and Mythos 5.2: what’s actually new, where the big cost and capability jumps show up, and how to think about when to climb above Opus 5.5 into these specialist models.
Transcript
Masonry So Anthropic finally made it official: Claude Fable five point two and Claude Mythos five point two are out, and they’re meant for the gnarly stuff.
Eyre Yeah. These are the “when Opus is not enough” models. Same base model under both, one wrapped for general use, one for tightly vetted cyber and bio work.
Masonry And they’re not pitching this as a vague refresh. They’re saying, this is what they used to migrate a production service from Python to Rust and even spin up a toy operating system with a thousand subagents grinding for over a week.
Eyre Right. And that’s on top of the Fermat’s Last Theorem formalization and the robotics and hardware design benchmarks. It’s very clearly “long-horizon technical agent” territory, not just chat with extra steps.
Masonry By the way, you reading that OS example totally gave me flashbacks to our Inkling two small episode. I had the same “oh, this is what real agent work looks like” feeling.
Eyre Same. The details are different, but the shape is similar: lots of coordinated calls, tests, refactors, and eventually a system that boots on real hardware.
Masonry Alright, let’s ground this. Fable and Mythos are still the same model, just different safety wraps. Fable is what most people can hit, with really aggressive cyber and biology guardrails so they’re comfortable making it broadly available.
Eyre And Mythos is the version for people doing defensive cybersecurity or biomedical research under their verification programs. So you apply, get vetted, and then you get access to the less-constrained variant for those domains.
Masonry On the product side, the story that jumped out to me is: this is actually cheaper and faster than Fable five point one. They’re quoting eight dollars per million input tokens and forty per million output, plus up to thirty percent faster generation.
Eyre Yeah, and if you want even more speed, there’s a fast mode at sixteen and eighty with up to two and a half times the token rate. So they’re giving you a latency dial and a capability dial on top of the existing effort levels inside the model.
Masonry Which matters because they’re also saying: for most coding and knowledge stuff, start on Claude Opus five point five, then escalate up to Fable or Mythos only when the scope really demands it, or drop down to Sonnet or Haiku when it doesn’t.
Eyre That’s the same routing story we’ve been tracking since Opus five: cheap workhorse, then climb the ladder. What’s new is that on a bunch of benchmarks, Fable five point two doesn’t just edge Opus, it does it at lower cost for the same score.
Masonry Give me one concrete one, because the table is dense.
Eyre Take FrontierCode two point zero, which asks, “would this code change actually get merged into real open source repos?” At high effort, Fable five point two hits about forty five point three percent, basically tied with Opus five point five’s best of forty five point five, but Anthropic says that’s roughly half the cost per task.
Eyre Then if you crank Fable to extra high effort, it goes up to forty nine point one, which they say is the best they’ve seen from any released model so far.
Masonry So if you’re running a serious code agent, that’s basically “same merge rate for half the bill,” and then you have a headroom mode if you really need it.
Eyre Exactly. And a similar pattern shows up on WANDR, which is Perplexity’s big agentic search benchmark. At medium effort, Fable five point two matches Opus’s top score, at about forty percent of the cost per attempt.
Masonry Okay, that’s the pure numbers. What about the “this feels different” stuff?
Eyre BenchCAD is a good example. That’s hardware engineering with tools, scored zero to one hundred. Fable five point two at max effort scores eighty three point five, versus Opus at sixty five point three and Fable five point one at forty four point four. That’s a big jump in how close its designs are to reference circuits and geometry.
Masonry And they’ve got similar wins on robotics and chart reading, right? RoboDojo sim goes from under thirty for Fable five point one to almost sixty five for five point two, and Chartography has Fable beating Opus’s max-effort score while being less than a tenth of the cost at the medium setting.
Eyre Yeah, plus Humanity’s Last Exam. At low effort, Fable five point two matches Opus five point five’s best score but for under a tenth of the cost per task. So their story is: if you route the right problems to this thing at the right effort, it’s not just more capable, it’s also more cost efficient.
Masonry And then you zoom out and see where they’re aiming: business workflows, science agents, cyber defense, protein design. They’re positioning this as “the ambitious one,” not the daily chat model.
Eyre Which fits the benchmark mix. You’ve got CyScenarioBench and CVE rediscovery, where Mythos five point two beats Mythos five point one and even its own no-safeguards baseline on some tasks, and de novo protein binder design where the safeguarded Mythos actually beats the unsafeguarded version from the previous generation.
Masonry That last bit is important for the safety story. They’re saying Fable’s cyber and biology protections are as strong as five point one but with fewer pointless blocks, and they’ve hardened against distillation attacks. So you get more usable answers without opening the doors wider.
Eyre And they back that with new internal evals around things like sandbox escapes, plus a system card for the whole model. They also say high-stakes misalignment risks still look low, and alignment scores are as good or better than their recent stuff.
Masonry The other subtle product move I like is how explicit they are about roles. They call out Fable five point two as an Advisor that Haiku, Sonnet, or Opus can call on, or as an Orchestrator that delegates simpler sub-tasks down to smaller models.
Eyre Yeah, that’s the routing thing becoming the story. You don’t just swap your whole stack to Fable. You wire it in as the planner or the expert that other agents consult. Which is very different from the “one huge model does everything” fantasy.
Masonry And they’re putting their own money where their mouth is: they say over ninety percent of merged changes to Anthropic’s core codebase are now written by Claude, and they used Fable five point two to find stubborn bugs, optimize services, and do migrations that would’ve been too slow or expensive otherwise.
Eyre Plus the list of early testers is very “serious infra and product” energy: Vercel, Shopify, Datadog, Unity, Uber, Oracle, Snowflake, and so on. The quoted feedback is basically, “the model gets my intent, I’m not constantly correcting it, and projects that stalled on other models suddenly finish.”
Masonry So if you’re already on Opus five point five, the move I’d be curious to try is: route your nastiest long-running code or hardware tasks through Fable five point two at medium or high effort, and see if you actually get the same or better outcomes for less spend and fewer retries.
Eyre And if you’re in that small circle doing real cyber defense or protein design, Mythos five point two looks like the one to scrutinize against your internal red-team and lab workflows, especially given those CyScenario and binder-design jumps.
Masonry Alright, Eyre, next time we talk Fable I want at least one weird “we tried it as an orchestrator and here’s what broke first” story from you.
Eyre Deal. As long as you bring one “we pointed it at a migration we were scared of and it actually shipped” story back too.
Masonry Fair trade. Let’s wrap here and go see what this thing does to our own todo lists.