Anthropic's Fable/Mythos 5.2: The Model That Wrote an OS From Scratch and Solved Open Physics Problems

Anthropic ships its most capable models yet — claiming solved open physics problems, a Python-to-Rust migration, and a prototype OS built by 1,000 subagents. The real story is the pricing and the safety architecture splitting one model into two access tiers.

Anthropic has released Claude Fable 5.2 and Claude Mythos 5.2, which are the same underlying model split into two access tiers by safety guardrails. Fable ships with cybersecurity and biology safeguards for general availability; Mythos removes those restrictions for vetted researchers who apply through Anthropic's Life Sciences and Cyber Verification Programs. The pricing — $8 per million input tokens, $40 per million output — undercuts their previous Fable/Mythos generation while delivering 30% faster output. A fast mode doubles throughput at double the price.

The benchmark claims are staggering in places. Fable 5.2 scores 49.1% on FrontierCode 2.0 at extended effort (versus 45.5% for Opus 5.5 and 42.8% for GPT-6 Astra), 89.3% on OSWorld 2.1, and 74.2% on Humanity's Last Exam with tools. On the Remote Labor Index — a measure of how much real knowledge work the model can automate — Fable 5.2 hits 41.3% at high effort, roughly double Fable 5.1's 17.9%. Mythos 5.2 leads on cybersecurity, finding 10.2% more vulnerabilities than Opus 5.5 and 14.4% more than Mythos 5.1 on CVE Rediscovery. The headline demos tell the story Anthropic wants told. One team of agents migrated a critical production service from Python to Rust, yielding a 26× startup speed improvement. An engineer tasked Fable 5.2 with building a prototype operating system from an empty directory; the model coordinated over 1,000 subagents for eight days, producing a bootable system with drivers, file storage, a windowing system, login screen, and installer. The model also formally proved Fermat's Last Theorem in Lean in 30 hours and solved two open problems in black hole physics involving holographic duality — both now posted as preprints. The science results deserve particular scrutiny. On theoretical physics, Fable 5.2 completed a calculation laid out in 2003 regarding phase transitions in small universes under holographic duality, and found the exact bound of a 15-year-old technique for ruling out candidate quantum gravity theories. On protein folding, Mythos 5.2 devised a novel approach to predicting the full capsid structure of adeno-associated viruses by running overlapping chains through existing protein-folding models and merging results using known symmetry constraints. These are real, verifiable claims — but they're also cherry-picked showcases from prerelease testing, not peer-reviewed results. Anthropic reports that over 90% of code changes merged into its own core codebase are now written by Claude with human guidance. This is a remarkable internal metric, but it's also self-referential: the company building the tool is measuring its own use of the tool inside its own engineering culture. Whether that ratio holds outside Anthropic's workflows — with their presumably Claude-optimized tooling, prompting practices, and code review norms — is an entirely open question. The Fable/Mythos split is the most interesting structural decision here. One model, two access levels: the broad public gets guardrails that block most cybersecurity and biology tasks, while vetted organizations get unfiltered capability. This is a pragmatic approach to dual-use risk — better than either blanket restriction or blanket access — but it concentrates the gating decision with Anthropic. Who qualifies for Mythos access, and by what criteria, becomes a real governance question as capability grows. The competitive picture is tightening. On AutomationBench, Gemini 4 Carbon actually leads at 53.9% versus Fable's 51.4%. On several other benchmarks, the margins between Fable 5.2, GPT-6 Astra, and Gemini 4 Carbon are single digits. The era of clear model dominance appears to be compressing into benchmark-by-benchmark knife fights, which makes pricing, API ergonomics, and ecosystem integrations increasingly decisive.