Claude Fable 5.2 Tops Benchmarks, Beats Opus 5.5, Builds OS with 1000 Agents

Introduction

Anthropic released Claude Fable 5.2 and Mythos 5.2, targeting high-difficulty tasks like OS development, mathematical formalization, theoretical physics, protein structure prediction, hardware design, and robot control rather than general QA. Fable 5.2 scores 49.1% on FrontierCode 2.0, surpassing Opus 5.5 (45.5%), GPT-6 Astra (42.8%), and Gemini 4 Carbon (41.1%). It also achieves 65 on the Artificial Analysis Intelligence Index v5, topping that leaderboard. Notably, Fable 5.2 coordinated over 1,000 agents for 8 days to generate a bootable operating system from an empty folder. However, Anthropic recommends Opus 5.5 for daily coding and knowledge work, reserving Fable for complex, open-ended, long-horizon reasoning tasks, and Mythos for vetted high-risk domains like cybersecurity and life sciences.

Fable and Mythos: Same Base, Different Safety Boundaries

Fable 5.2 and Mythos 5.2 share the same underlying model but differ in safety restrictions. Fable is open to general users with added safeguards against cybersecurity and biological research misuse. Mythos retains more complete professional capabilities and is restricted to vetted security organizations and life-science institutions. The following table summarizes the positioning:

Opus 5.5 : Daily programming and knowledge work; regular access.

Fable 5.2 : Complex engineering, mathematics, research, and multi-agent tasks; available via Claude, cloud platforms, and API.

Mythos 5.2 : High-risk tasks in cybersecurity and life sciences; access requires qualification review.

This tiered approach is worth emulating: model capability should match task risk, budget, and permissions. Strong models handle planning, review, and complex judgment; low-cost models handle high-frequency execution; sensitive actions are gated by an independent policy layer.

FrontierCode and FrontierSWE: Leading Complex Software Engineering

On FrontierCode 2.0, Fable 5.2 scores 49.1%, up 16.7 percentage points from Fable 5.1's 32.4%. Opus 5.5 scores 45.5%, GPT-6 Astra 42.8%, Gemini 4 Carbon 41.1%. On FrontierSWE v2, Fable 5.2 reaches 69.9%, a 13.6-point gain over its predecessor. At lower reasoning intensity, Fable 5.2 approaches Opus 5.5's peak performance at roughly half the per-task cost. Anthropic also disclosed that over 90% of merged code changes in its core codebase were written by Claude under human guidance, though this measures contribution volume, not a 90% engineer efficiency gain; test pass rates, rework, rollback rates, and final delivery quality must also be considered.

1000 Agents Build an OS in 8 Days

An engineer tasked Fable 5.2 with designing an operating system from scratch. The model autonomously coordinated over 1,000 sub-agents for 8 continuous days, handling code writing, testing, and review. The resulting system boots on a laptop and includes a custom kernel, disk, keyboard and network drivers, file storage, windowing interface, login screen, installer, and built-in Claude functionality. The challenge lies not in individual agents writing code but in task decomposition, state sharing, change merging, failure handling, and architectural consistency. Enterprises wishing to replicate this workflow need task trees, isolated branches, permission isolation, checkpoints, and unified regression testing. Fable 5.2 also migrated a critical production service from Python to Rust, achieving a 26× startup speedup; migration tasks demand attention to behavioral consistency, performance metrics, dependency changes, rollback paths, and production validation.

30 Hours to Formalize Fermat's Last Theorem

While the mathematical proof of Fermat's Last Theorem is long established, translating it into Lean for computer verification remains a complex engineering effort. Anthropic previously used an unreleased research model with human researchers over 11 days. This time, a team of 1,000 Fable 5.2 agents completed the formalization in 30 hours, reducing proof length to one-seventh and token consumption to one-third of the prior effort. Under the same configuration, Opus 5.5 achieved only ~27% completion in 30 hours. The key value is the proof checker's explicit feedback: compilation failures, type errors, and proof gaps become the basis for the next repair iteration. The model proposes and organizes; the independent verifier (Lean) judges true completion.

Black Holes, Proteins, and Electronic Devices

In theoretical physics, Fable 5.2 solved two open problems related to black holes and holographic duality: a phase-transition mechanism during black hole formation and the applicability boundary of an analytical method for quantum gravity. Results require scientific review, independent replication, and peer review. Mythos 5.2 combined two open-source protein folding models to predict gene-therapy vector capsid structures; 28 of 34 test capsids met the accuracy threshold. In hardware experiments, Fable 5.2 designed a home voice-assistant hub, a USB conference speaker, and a handheld electronic pet from a single prompt, producing product specs, component selection, four-layer PCB layouts, 3D-printable enclosure models, and accompanying software. All three boards reached manufacturable state in under a day and functioned without electrical design changes after fabrication. The article emphasizes that model output is a design proposal; production still requires material validation, reliability testing, regulatory review, and hardware engineer sign-off.

Robot Control: From Code to Physical Operation

On the RoboDojo Sim dual-arm robot simulation benchmark, Fable 5.2 achieves a 64.9% success rate, far above Fable 5.1's 27.6% and GPT-6 Astra's 29.8%. Official demos show robots unscrewing bottle caps, solving Rubik's cubes, and drawing robot figures. Physical agent evaluation must go beyond final success rate to record action counts, collision risk, visual latency, retry attempts, hardware wear, and human interventions. The model can command actions, but safety boundaries should remain the responsibility of the robot controller and hardware systems.

Mythos Cybersecurity Capabilities and Limits

Mythos 5.2 rediscovers 89.6% of known vulnerabilities in the internal CVE Rediscovery test, up from 78.3% for Mythos 5.1. In a sandbox-escape suite of over 1,800 scenarios, the rate of using external credentials or files drops to 3.1% from 21%. As capabilities rise, Anthropic deliberately caps performance in AI accelerator kernel optimization and LLM architecture improvement to Opus 5.5 levels. This illustrates that model release involves capability boundary design: networks, credentials, exploits, training code, and production systems should be governed by project-side policy gateways, not left to the model's discretion.

Pricing, Fast Mode, and Usage Routing

Fable 5.2 is available via Claude, AWS, Google Cloud, and Microsoft Azure using the model identifier claude-fable-5-2. Standard API pricing is reduced ~20% from the previous generation: $8 per million input tokens, $40 per million output tokens. Fast Mode increases token generation speed up to 2.5× at double the price: $16/M input, $80/M output. Anthropic reiterates its routing guidance: Opus 5.5 for daily coding and knowledge work; Fable for complex engineering, mathematics, and research; Mythos for vetted high-risk cybersecurity and life-science tasks. Domestic developers integrating tools like Claude Code with compatible APIs can refer to Code80 (https://code.ai80.vip/home) for endpoint configuration, while retaining project-side control over model permissions, keys, data retention, budgets, and production audit logs.

FAQ

Is Fable 5.2 suitable for all development tasks?

No. Daily programming and knowledge work should use Opus 5.5; Fable is for complex engineering, mathematics, research, and long-horizon agent tasks.

What is the difference between Fable and Mythos?

Fable targets general users with added cybersecurity and biological research restrictions; Mythos retains fuller professional capabilities for vetted institutions.

Can 1,000 agents really collaborate stably?

This is an Anthropic demonstration. Enterprise replication requires task decomposition, isolated sandboxes, state management, budgeting, rollback, and human approval—not merely adding more agents.

Can research examples be treated as paper conclusions?

No. Black-hole physics, mathematical formalization, and protein structure results still need domain expert review, independent verification, and peer review.

How does Fable 5.2 reduce enterprise model costs?

Route high-frequency routine tasks to Opus, complex tasks to Fable, and vetted high-risk tasks to Mythos; track tokens, retries, human rework, and task success rates.

How can domestic developers access compatible APIs?

Use Code80 for integration guidance while managing model permissions, keys, data retention, call budgets, and audit logs internally.