Fable 5.2 Is the Best AI Ever. So Why Am I Not Switching Yet?

The host introduces Anthropic’s new Fable 5.2 flagship model and Mythos 5.2, released as the company approaches its IPO. Fable leads GPT-6 Astra by four points on Frontier Code and six points on Terminal Bench Science 0.1; Mythos leads prior cybersecurity results by about ten percentage points. Fable scored 64.9% on Robo Dojo Sim, 25 points above Opus 5.5, while Mythos reached 85.1% in protein design. On the Remote Labor Index, Fable scored 41.3%, nearly double Opus 5.5’s 21% and far above the 2.5% recorded by the best agent a year earlier.

Fable 5.2 leads the Artificial Analysis Intelligence Index with a score of 65, ahead of Gemini Carbon, Opus 5.5, and GPT-6 Astra, although gains were smaller in knowledge work and terminal coding. Anthropic cut API prices 20%, to $8 per million input tokens and $40 per million output tokens. At maximum effort, Fable costs $26.54 per benchmark task—over twice Fable 5.1 and 2.6 times Opus 5.5. Extra-high effort reduces that to $14.83, while high costs $9.08. Anthropic recommends medium for normal use and extra-high only for long-running agentic work.

Anthropic says 90% of its code changes are written by Claude under human supervision, with Fable 5.2 intended for stubborn bugs, performance optimization, and high-value projects too slow or costly for other models. Its example was migrating a critical production service from Python to Rust. In an eight-day experiment, Fable coordinated more than 1,000 subagents to build a bootable operating system with drivers, storage, windows, login, installation, and quad functionality. Anthropic also reported two theoretical-physics solutions, accurate modeling of 28 of 34 gene-therapy capsids, and a machine-readable formalization of Fermat’s Last Theorem.

Community testing produced strong but uneven demonstrations. Weander Builds said medium-effort Fable found more than ten severe vulnerabilities that Opus 5.5 had created. In Pavel Huren’s benchmark with 105 planted bugs across two repositories, Fable found 58, compared with 51 for Sonnet 5.5 Max and roughly 40–45 for most major models, using fewer turns and less money. Fable’s 3D castle rendering was much more detailed than Opus 5.5, and Riley Brown produced a polished 60-second advertisement with detailed zooms, lighting, and physics. Allie K. Miller said it handled ambitious tasks—including a 3,100-city water-safety map and a monetizable YouTube channel—better than any model she had tested.

The Every team described Fable 5.2 as a major shift from Opus 5.5: instead of keeping a professional in the loop, users can hand it an entire product and let it build, test, and refine for hours or days. Their example combined Adobe Premiere-style editing, After Effects-style titles and effects, and Ableton-like music mixing into a working video editor. The host connects this to the older “SaaS apocalypse” argument: although maintaining custom software once seemed too expensive and distracting, Fable may make personal professional software practical. The team also warned that its strengths are spiky: extraordinary on certain ambitious tasks, ordinary on many others.

Ethan Mollick framed Fable 5.2 as “Jagged AGI.” He found it capable of producing a 182-page paper, an accurate course, and elaborate rule systems, but the course was joyless, the paper repetitive, and the project buried under bureaucracy. Critics and fact-checkers corrected errors but did not recognize that the work was too long, dull, or overcomplicated. Mollick argues that knowing when one’s own work is not good enough is central to expertise, and Fable does not reliably do that; he would not trust it to run a business independently, even though its plans could impress a team of MBA students.

NLW asked Fable to create an impressive demonstration. In the standard interface it chose technically checkable but emotionally obscure feats: a sentence counting its letters, self-reversing harmonized music, and a QR code built from scratch. In a coding harness, it made a more compelling three-minute browser-generated film whose visuals, score, narration, and lettering were produced largely from code, closely resembling Mollick’s own result from a similar prompt. Asked for an AI Daily Brief project, it proposed tracking announcement-to-availability dates for models—a factual gap but, in the host’s view, an unfathomably boring idea. Its better backup ideas involved employer AI statements and grading AI predictions.

NLW’s working thesis is that Fable 5.2 is “opportunity AI,” not merely “efficiency AI”: its value comes from enabling projects people previously would not attempt. He highlights evidence that it can build games with functioning mechanics rather than attractive but unplayable prototypes. Anthropic’s Shulta Douglas called it the first model that met his standard for “seriously fun and beautiful games,” while Matt Schumer had it create a playable Harry Potter-themed Fortnite-style battle royale that other people were using. NLW suggests this may change the question from whether AI should make a deck or website to whether a game is the better interface for an idea.

Several reviewers said Fable 5.2 is often more annoying than Opus 5.5 because it is expensive, token-hungry, stubborn, and burdened by guardrails. Daisy Holman’s suggested division is to use Opus for specific tasks with a known desired outcome and Fable for open-ended creative work, though NLW had not fully confirmed that pattern. Anthropic’s Alex Albert recommends asking Fable to criticize an idea before building it, specifying where approval is required, starting at medium or high effort, reserving extra-high for difficult problems, and assigning a whole project rather than a next step. For coding, Anthropic says to tell Fable that it owns the codebase; that instruction makes it more likely to merge duplication and fix related problems instead of merely reporting them.