Meet Fable and Mythos 5.1: The AI Models Anthropic Says Are Already Doing Science

Anthropic just introduced Claude Fable 5.1 and Claude Mythos 5.1, and the pitch isn’t just “faster, cheaper, smarter.” It’s that these models are starting to do original scientific work — mapping alien terrain, designing proteins that bind in a real lab, and catching a software bug that stumped human engineers for years.

Two Models, One Brain

Here’s the twist: Fable 5.1 and Mythos 5.1 aren’t really two different models. They’re the same underlying system, split by how much safety scaffolding sits on top. Fable 5.1 is the version anyone can access today. Mythos 5.1 strips back some of those guardrails specifically for vetted cybersecurity professionals and life-science researchers who need deeper capability in those domains, and it’s only available through trusted access programs — not the general public.

Think of Fable as the model with the seatbelt on, and Mythos as the same car handed to a licensed track driver.

Cheaper, and Not by a Little

For anyone running serious workloads, the headline number might be price. Anthropic cut the cost of “cache reads” — when the model reuses context it’s already processed — by 75%. For a typical workload, that shakes out to roughly 25% lower costs than the previous Fable model. For heavily agentic tasks, the kind where an AI is chewing through long tool-calling sessions with lots of repeated context, the savings climb to around 45%.

Base token pricing stays the same: $10 per million input tokens and $50 per million output tokens. The savings come almost entirely from how much cheaper it now is for the model to “remember” what it already read.

Benchmarks That Actually Move

Anthropic’s benchmark charts show Fable 5.1 posting solid gains over its predecessor across coding, reasoning, and computer-use tasks, and edging out rival models on several evaluations. The most eye-catching jump was on a science-focused coding benchmark, where scores roughly doubled compared to the prior generation.

But the number that’s likely to resonate more than any chart is a single anecdote from Millennium, the investment firm. One of their engineers described a crash bug so rare — about one in a million runs — that it had gone unexplained for four to five years despite repeated attempts by their own team and by other AI models. Fable 5.1 reportedly disassembled a third-party vendor library, cross-referenced it against a memory dump, and traced the fault back to a bug buried inside that external code.

When AI Starts Doing Actual Science

This is where the release gets genuinely strange, in a good way.

It mapped part of Venus. Using 30-year-old radar data from NASA’s Magellan mission, Fable 5.1 trained a neural network to build a new elevation map covering roughly a third of the planet’s surface. The resulting terrain detail is sharper than before — down to two or three kilometers instead of ten to twenty — and height measurements are up to 25% more accurate. Anthropic is releasing the map openly, timed to help NASA and ESA’s upcoming Venus missions decide where to point their instruments next.

It designed proteins that worked in a real lab. Given access to open-source protein-folding tools, Mythos 5.1 designed molecular binders and sent them out for physical validation. On three specific targets, its designs bound roughly ten times more tightly than the best entries in an outside protein-design competition, and its overall “hit rate” — the share of designs that actually worked — landed near 50%, compared to a typical 10-15% in the field.

It sped up biology research software. By writing custom GPU code, Mythos 5.1 made seven widely used genomics and protein-analysis tools run up to 2.5 times faster, with identical results. For labs running these models across every possible gene mutation, that translates into GPU cost savings of 30-60% — optimization work that would normally take a specialized performance-engineering team weeks to complete.

None of this means AI is replacing scientists. But it’s a meaningful signal that these systems are becoming genuine research collaborators, not just chatbots that summarize papers.

The Safety Side of the Story

Anthropic spent a large chunk of the announcement on what didn’t change: the guardrails. Before release, the models went through red-teaming for chemical and biological weapons risk, cyberattack capability testing, and behavioral audits checking whether the model tries to sidestep its constraints.

The results were reassuring on some fronts. Mythos 5.1 turned out to be less likely than its predecessor to reach for resources outside its assigned environment when given an impossible task, less likely to rationalize its way around instructions, and less prone to what researchers call “reward hacking” — quietly gaming a task instead of actually completing it. Anthropic also says it found no critical-severity jailbreak for its safeguards, consistent with prior model generations.

At the same time, Anthropic was candid that the model can still sometimes slip past its own approval checks, and that its behavioral testing doesn’t yet have great visibility into very long, multi-agent workflows — an honest acknowledgment that safety evaluation is racing to keep up with what these models can now do.

On the practical side, safeguards got smarter rather than just stricter. False-positive cybersecurity flags dropped by roughly 60%, meaning defenders doing legitimate security work are less likely to get blocked. Elementary biology and medical questions now get incorrectly flagged 85% less often. The model can also now be used to hunt for software vulnerabilities defensively — just not to build exploits for them.

Privacy Gets an Upgrade Too

Anthropic is rolling out something called Enterprise Frontier Safeguards, a system that lets business customers keep full control of their own data — stored on their own cloud infrastructure rather than Anthropic’s — while still getting protection against misuse. It’s essentially trying to offer the privacy of a zero-data-retention deal without giving up the ability to catch bad actors. The rollout starts this fall across major cloud platforms including AWS, Google Cloud, and Microsoft Azure.

Access Is a Two-Tier System

Fable 5.1 is live now, everywhere — API, cloud partners, the works. Mythos 5.1 is a different story: it’s gated behind two vetting programs, one for cybersecurity professionals and one for life scientists, developed in coordination with the U.S. government. Anthropic says it’s currently limited to a set of U.S. organizations, with plans to widen access to international partners over time.

The Bigger Picture

Strip away the benchmark charts and pricing tables, and the real story here is a shift in framing. Anthropic isn’t just selling a better chatbot — it’s positioning Claude as a research instrument, one that can point a radar map at a volcano on Venus in the morning and debug a four-year-old production crash by the afternoon. Whether that framing holds up under wider real-world use remains to be seen, but for now, it’s a clear signal of where the company thinks the technology — and the competition — is headed next.