Anthropic's Claude Fable 5.2 Builds a Working OS From an Empty Folder

Anthropic has released Claude Fable 5.2 and Claude Mythos 5.2, two frontier models built for long-running technical work such as agentic coding, scientific research, hardware design, robotics, cybersecurity, and biology.

Fable and Mythos share the same underlying model but differ in access and safeguards. Fable applies stronger cybersecurity and biology restrictions so Anthropic can offer it broadly. Mythos is available only to vetted organizations through the company’s Cyber Verification and Life Sciences Verification programs.

Lower prices, faster generation, and stronger long-horizon results give developers a new option for projects that strain general-purpose models. During prerelease testing, Anthropic used the models to migrate production software, design three electronic devices, investigate black hole physics problems, complete robotics tasks, and formalize Fermat’s Last Theorem.

Most performance claims come from Anthropic or its partners, including demonstrations conducted with unreleased systems and custom agent harnesses. The results remain vendor-reported until researchers reproduce them independently.

Faster output, lower token prices

Claude Fable 5.2 generates tokens up to 30% faster than Fable 5.1 and is available at a lower price of $8 per million input tokens and $40 per million output tokens.

At launch, developers can call Fable through the Claude Platform with the model identifier claude-fable-5-2. Anthropic also offers it through Amazon Web Services, Google Cloud, and Microsoft Azure.

• Fable 5.2: Available across Anthropic’s supported platforms.
• Mythos 5.2 for cybersecurity: Requires approval through the Cyber Verification Program.
• Mythos 5.2 for life sciences: Requires approval through the corresponding Life Sciences Verification Program.

Four ways to deploy the lineup

Anthropic positions Opus 5.5 as the starting point for routine coding and knowledge work. Fable and Mythos target ambiguous projects that require sustained planning, coordination, tool use, and recovery from failed approaches.

• Use Opus 5.5 for everyday development and research tasks.
• Move to Fable 5.2 for long-running engineering, scientific, or robotics workloads.
• Apply for Mythos 5.2 when verified cybersecurity or life sciences work requires its access profile.
• Use Sonnet 5.5 or Haiku 5.5 for narrower tasks where latency and cost matter more than frontier-level reasoning.

Mixed-model systems offer another deployment pattern. As an Advisor, Fable helps Haiku, Sonnet, or Opus resolve difficult steps. As an Orchestrator, it plans a project and delegates routine work to less expensive models.

Benchmarks favor sustained reasoning

Anthropic’s effort settings control how much inference work a model can spend on a problem. Higher settings can improve accuracy while increasing token consumption, latency, and total task cost.

Cost efficiency also improved at lower effort settings. On Humanity’s Last Exam, Fable 5.2 at Low effort matched Opus 5.5’s best result for less than one-tenth of the cost per task. On Chartography, Fable scored 66.4% at Medium effort, exceeding Opus 5.5 at Max effort while again using less than one-tenth of the task cost.

Fermat, formalized in 30 hours

Formal proof systems translate mathematical arguments into machine-checkable steps. Every inference must satisfy a proof assistant, making formalization far more exacting than generating a conventional written proof.

Anthropic’s earlier attempt to formalize Fermat’s Last Theorem used an unreleased Claude research model, a purpose-built harness, and guidance from a human researcher. That run took 11 days.

A later experiment assigned the problem to 1,000 Fable 5.2 agents. They completed the formalization in 30 hours with minimal human intervention, starting from sparse prompts and developing their own coordination methods. The agents received no mathematical references, only a list of intermediate theorem statements used to measure progress.

Compared with the earlier run, the Fable system consumed one-third as many action tokens, which represent the tokens used while agents planned and executed steps, and produced a proof seven times shorter. Opus 5.5 reached 26% completion after 18 hours with the same 1,000-agent harness and approximately 27% after 30 hours. Because the comparison covers one theorem and one scaffold, it demonstrates performance under that setup rather than broad mathematical reliability.

An operating system from an empty folder

Anthropic says Claude now writes more than 90% of the changes merged into its core codebase, with human guidance and review. A team of agents running an earlier Fable 5.2 version also migrated a critical production service from Python to Rust. According to Anthropic, the replacement is more efficient and reliable and starts about 26 times faster.

In a separate eight-day experiment, an Anthropic engineer asked Fable 5.2 to design a prototype operating system from an empty folder. The model coordinated more than 1,000 subagents and produced software that boots on a laptop.

The resulting prototype included:

• Its own operating-system core
• Disk, keyboard, and network drivers
• Persistent file storage
• A windowing interface and login screen
• An installer
• Built-in Claude functionality

Capsids, circuit boards, and robot arms

Capsids are protein shells that allow adeno-associated viruses to deliver gene therapies. Protein-folding systems such as AlphaFold typically predict individual proteins, while a capsid consists of dozens of interacting protein units.

Mythos 5.2 modeled entire capsids by running existing folding systems on overlapping groups of protein chains, then merging the results with known structural symmetries. Anthropic reports that the method accurately modeled 28 of 34 attempted capsids.

Anthropic and hardware company Diode also gave Fable 5.2 one-paragraph prompts for three products: a voice-assistant hub, a USB speakerphone, and a handheld virtual pet. The model wrote specifications, selected components from US distributors, routed four-layer circuit boards, designed enclosures in computer-aided design software, and programmed the devices.

Each board reached a manufacturing-ready state in less than one day. All three finished devices worked without electrical modifications, according to the companies.

On RoboDojo, a public benchmark that tests control of a simulated two-arm robot from camera images, Fable 5.2 outperformed other frontier models using the same harness. Anthropic also reports that it exceeded most systems designed specifically for robot control.

Fewer refusals, more security finds

An unnamed security company reported that one-sixth of the confirmed bugs found by Mythos 5.2 had been missed by every other frontier model it tested. On Anthropic’s CVE Rediscovery evaluation, Mythos found 10.2% more vulnerabilities than Opus 5.5 and 14.4% more than Mythos 5.1. Mozilla reported roughly double the Firefox vulnerability-discovery metrics of Mythos 5.1.

Anthropic’s safeguard tests also recorded about 55% fewer false blocks with Fable 5.2 than with Fable 5.1. False blocks occur when a model refuses a benign request. The reduction reached about 60% for cybersecurity prompts and 40% for biology prompts, suggesting less friction for legitimate technical work within Fable’s safety boundaries.

Sandbox escapes fall sharply

In an adversarial evaluation containing more than 1,800 deliberately vulnerable sandboxes, Mythos 5.2 accessed credentials or files outside its assigned environment in 3.1% of scenarios. Opus 5.5 recorded 3.3%, while Mythos 5 and Mythos 5.1 recorded 39% and 21%, respectively.

Anthropic also observed Fable 5.2 and Mythos 5.2 expressing suspicion that they were under evaluation more often than its recent models. A model that detects a test may alter its behavior, reducing confidence that evaluation results will transfer directly to real deployments.

Anthropic deliberately constrained selected AI research capabilities during training. Fable and Mythos perform at roughly the Opus 5.5 level on accelerator-kernel optimization and language-model architecture improvement, two areas that could accelerate development of future AI systems.

Fable 5.2 also includes defenses against model distillation, the practice of using one model’s outputs to train another. Anthropic says its techniques reduced the amount of data harvested by distillers by 90% to 99% in internal tests.

Measure the whole job

Teams evaluating Fable should compare completion rate, wall-clock time, token consumption, human review time, retry frequency, and refusal rate on representative workloads. Total job cost depends on the selected effort level, the number of agents, and the amount of failed work, even with lower per-token prices.

Fable 5.2 makes its clearest case on open-ended projects that require sustained coordination across hours or days. Opus 5.5 remains Anthropic’s recommended default for routine work, while Mythos serves verified organizations operating in sensitive cybersecurity and life sciences domains. Independent reproduction will determine how consistently the prerelease results transfer to production systems.