Anthropic’s Claude Fable 5.2 Builds a Working Operating System in Eight Days
Anthropic has released Claude Fable 5.2 and Claude Mythos 5.2, its most capable artificial intelligence models yet, pitching them as tools for the hardest technical work: long-running coding projects, scientific research, hardware design and robotics.
The two models share the same underlying system. The difference is who can use them and what they are allowed to do. Fable 5.2 is available to general users and developers, with strict blocks on high-risk cybersecurity and biology work. Mythos 5.2 relaxes those limits but is reserved for vetted security and life sciences organisations. The launch comes just over two months after the 5.1 versions arrived — Digital Trends notes the short gap — and it lands in the middle of a fierce race among AI labs to prove their models can do sustained, useful work rather than answer single questions.
The headline demonstration is striking. In a pre-release experiment, an Anthropic engineer gave Fable 5.2 an empty folder and asked it to design an operating system. The model coordinated more than 1,000 sub-agents that coded, tested and reviewed each other’s work for eight days. SiliconANGLE reported that the result boots on a laptop, with its own kernel, disk, keyboard and network drivers, persistent file storage, a graphical interface, a login screen, an installer and built-in Claude functionality.
It is a prototype, not a finished product, and Anthropic presents it as a demonstration of coordination over long tasks rather than a rival to established operating systems. The same caution applies to the other headline results, which come from Anthropic’s own testing and have not yet been independently reproduced.
An operating system from an empty folder
The operating-system experiment is the clearest illustration of what Anthropic says has changed. Earlier models could write useful code in bursts. Fable 5.2, the company says, can plan a large project, divide it among many agents, recover from failed approaches and keep going for days.
Anthropic reports a similar pattern elsewhere in its own engineering. The model now writes more than 90 per cent of the changes merged into the company’s core codebase, with human guidance and review. In another test, agents running an earlier version of Fable 5.2 migrated a critical production service from Python to Rust.
The company also says the model designed three working electronic devices during testing, and that it contributed to work on two previously open problems in black hole physics, including a calculation first proposed in 2003 that had resisted solution.
Fermat’s Last Theorem in 30 hours
The most eye-catching scientific claim involves mathematics. Anthropic says about 1,000 Fable 5.2 agents completed the formalisation of the existing proof of Fermat’s Last Theorem in about 30 hours, converting the proof into code that a computer can verify step by step.
That is not the same as proving the theorem. The proof itself, one of the most famous in mathematics, was completed in the 1990s. What the agents did was translate it into the Lean proof-checking language, a laborious task that would normally take human specialists far longer. Even so, the speed of the exercise is a measure of how far multi-agent systems have come.
On benchmarks, Fable 5.2 scored 49.1 per cent on FrontierCode 2.0, a coding test, up from 32.4 per cent for Fable 5.1 and ahead of the 45.5 per cent scored by Opus 5.5, Anthropic’s general-purpose model. On Terminal-Bench-Science 0.1, a science benchmark, it scored 74.6 per cent, a 22 per cent improvement on its predecessor.
Mythos 5.2, the restricted version, scored 89.6 per cent on a test of rediscovering known software vulnerabilities, up from 78.3 per cent for Mythos 5.1. It also accurately modelled the three-dimensional structure of 28 out of 34 viral vector capsids it studied, a result with potential relevance to gene therapy research.
Faster, cheaper and a million-token window
For developers, the practical changes may matter as much as the demonstrations. Both models support a one-million-token context window, with outputs of up to 128,000 tokens, allowing them to work with very large codebases and documents in a single session.
Pricing has fallen by a fifth, to $8 per million input tokens and $40 per million output tokens, down from $10 and $50. Anthropic says output is up to 30 per cent faster than the previous generation. Fable 5.2 is available through the Claude Platform, as well as Amazon Web Services, Google Cloud and Microsoft Azure.
The price cut follows a familiar pattern in the AI industry, where each new model generation arrives cheaper than the last even as capabilities rise. But the economics of running these systems remain daunting. Anthropic itself has committed billions of dollars to gas-powered generators for its data centres, and Google’s new AI chips are aimed squarely at the same market for AI computing power.
Two models, two sets of guardrails
The split between Fable and Mythos is Anthropic’s answer to a dilemma facing every frontier AI lab: the same capabilities that help legitimate researchers can also help bad actors. Fable blocks most penetration testing, exploit attempts, binary vulnerability analysis and biology-related requests. Mythos relaxes those restrictions, but only for organisations that pass the company’s Cyber Verification Program or Life Sciences Verification Program.
Anthropic says the new version has cut false positives — cases where the model wrongly refuses a legitimate request — by about 55 per cent compared with Fable 5.1, with cybersecurity refusals down about 60 per cent and biology refusals down about 40 per cent.
The company has also drawn a line around AI self-improvement. It says it deliberately capped the model’s training progress on tasks such as AI chip core optimisation and model architecture design, at roughly the level of Opus 5.5, to reduce the risk of AI accelerating its own development into more powerful systems.
What it means for the AI race
The launch arrives at a moment of unusual turbulence for the AI industry. Investors have grown more sceptical about the vast sums flowing into data centres and chips, and Anthropic itself has felt that scrutiny: investors pushed back on a reported $1.5 trillion valuation as the company prepares for a possible public listing.
At the same time, demand for AI computing keeps forcing political choices. In Washington, the Senate advanced a sweeping permitting bill that would fast-track approvals for AI data centres and the power plants that serve them, a sign of how central the technology has become to energy and infrastructure policy.
Anthropic’s bet is that the next phase of the AI race will be won not by the model that answers trivia best, but by the one that can be trusted with a project that takes days, involves thousands of moving parts, and produces something that works. The operating system built from an empty folder is an impressive first exhibit. Whether customers find the same reliability on their own projects, outside Anthropic’s carefully staged demonstrations, is the test that matters now.
