Anthropic launches Claude Fable 5.2: "almost AGI" performance, but the model can't spot its own lapses in judgment

<p>Anthropic has unveiled Claude Fable 5.2 and Claude Mythos 5.2, two versions of the same model built for the most demanding technical work. As with earlier releases, Fable ships with filters that block risky uses in cybersecurity and biology and is open to paying subscribers, with some exceptions. Mythos has no such restrictions and is limited to organizations admitted to Anthropic's vetting programs.</p><h2>Faster, and strongest on long tasks</h2><p>According to Anthropic, Fable 5.2 responds up to 30% faster than version 5.1 and performs best on long-running tasks where it has to plan the work and coordinate other agents. A team of one thousand Fable 5.2 agents formalized the proof of Fermat's Last Theorem in Lean, a language whose proofs a computer can check step by step, in just 30 hours. With the same setup and the same amount of time, Opus 5.5 got through only 27% of the work.</p><p>Its widest lead over competitors is in robotics. On RoboDojo, a benchmark built around a simulated two-arm robot, Fable 5.2 scores 64.9% against 29.8% for GPT-6 Astra. Even though it is a general-purpose model, an Anthropic video shows it solving a Rubik's cube with robotic arms. The gap is smaller in coding: 49.1% on FrontierCode 2.0 versus GPT-6 Astra's 42.8%. At the High effort level, which allows longer reasoning, it matches Opus 5.5's best score at roughly half the cost per task, and it built a new operating system in eight days starting from an empty folder. In business workflow automation, Gemini 4 Carbon still comes out ahead (53.9% vs 51.4%). For everyday work, Anthropic itself recommends starting with Opus 5.5.</p><h2>Safety, filters and test awareness</h2><p>Before launch, the model was also reviewed by outside evaluators, including the UK AI Security Institute and METR. Anthropic rates the risk of misaligned behavior in high-stakes situations as low. In sandbox escape tests, which probe the isolated environment the model runs in during training and internal testing, Mythos 5.2 exploited deliberately planted flaws in just 3.1% of cases, down from 21% for Mythos 5.1. Fable still runs under classifiers, automated filters that block things like exploit writing, penetration testing and most biological research. Blocked requests are handed to less capable models: Opus 4.8 for cybersecurity and Opus 5 for biology. Fable 5.1 often blocked harmless questions too. In early tests of Fable 5.2, blocks fell by about 55% overall: 60% in cybersecurity and 40% in biology.</p><p>The thorniest issue is evaluation awareness, meaning the model's ability to tell when it is being tested. Mythos 5.2 suspected it might be under evaluation in up to 43% of conversations, more than Mythos 5.1 (35%) and Opus 5.5 (36%). In a software supply chain test run by the AI Security Institute, it raised the possibility of an alignment test in 96% of cases. A model that knows it is being watched may behave better than it would in real use, and Anthropic admits this weakens the conclusions of its own evaluations.</p><h2>Ethan Mollick's "Jagged AGI"</h2><p>Ethan Mollick, an entrepreneurship professor who studies AI's impact on work, tested Fable 5.2 early and asked whether it already counts as AGI. He uses a narrow definition: an AI as good as most human experts at intellectual tasks done on a computer. His answer was "almost." He calls it a "Jagged AGI": expert-level in many fields, but with gaps that are hard to predict. With no instructions available, the model turned a small USB keypad into a console with six games. It also wrote an economics paper that Mollick considers more advanced than what you would expect from a second-year PhD student.</p><p>Mollick notes that the mistakes are no longer made-up facts or bad math. They are now judgment failures much like those of human professionals and organizations. A website about W. H. Auden's 1941 literature course was accurate but boring. A research report grew to 182 pages because every critique triggered new analysis and nothing was ever cut. In a 3D-worlds project run by dozens of agent teams, the system wrote itself an 87,500-word internal rulebook: it had built 328 worlds in its first two days, but by the end publishing a single new one took about a day. The model caught none of these problems on its own. Critic and verifier agents fixed errors, but none of them asked whether the work had become too long, repetitive or unpleasant. That is why Mollick would not yet let it run a business on its own, even though he says Fable's business plan would have impressed him coming from a group of his MBA students.</p><h2>Availability and pricing</h2><p>Fable 5.2 is available on paid plans, on the Claude Platform, and through Amazon, Google and Microsoft's clouds. API pricing is $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for the previous version. During launch week, subscribers get one to three Fable-specific usage limit resets, depending on how long they have been subscribed. On Pro plans and standard Team seats, Fable 5.2 can use at most 50% of the weekly limits.</p>