Anthropic's Fable 5.2 and Mythos 5.2 top nearly every benchmark ever run, solve open science problems and build an operating system from scratch. Yet the most common early reaction is uncertainty about when to reach for it over Opus 5.5, which raises a bigger question about what frontier capability is now actually for.
The best model ever is a spiky one, and it only pays off if you get more ambitious.
Fable 5.2 is easy to call the most powerful model that has ever existed. It leads in coding, science, cyber and robotics, and it nearly doubles the best score on real-world contractor tasks. It is also expensive, token-hungry and, for most everyday work, no better than Opus 5.5. Its failures are subtle ones of taste and judgment: it doesn't know what humans find interesting, and it doesn't notice when its own work is a slog. The upshot is that this is opportunity AI, not efficiency AI. It is a model for whole projects and for things that weren't possible before, like complete software, live agentic video and playable games. And at this capability level, figuring out what a new model is really for takes weeks, not hours.
Fable 5.2 leads the field on the main benchmarks. On Frontier Code it is four points ahead of GPT-6 Astra and three and a half ahead of Opus 5.5. On Terminal Bench Science it is a full six points ahead of GPT-6 Astra. On cybersecurity, Mythos 5.2 beats the previous leader by about 10 percentage points in each case.
Anthropic published non-standard benchmarks for emerging capabilities. Fable 5.2 scored 64.9% on RoboDojo Sim, 25 points above Opus 5.5, whose own score was already a breakout result. Mythos 5.2 hit 85.1% on protein design, six points above Mythos 5.1.
On one agentic knowledge-work benchmark, Fable 5.2 completed about five points more tasks than Fable 5.1, Opus 5.5 and Gemini 4 Carbon. But its ELO score was lower than Opus 5.5's. It finishes more work without necessarily doing that work better than its smaller stablemate.
On the Remote Labor Index, which uses real contractor projects, Fable 5.2 scored 41.3%. That nearly doubles Opus 5.5's 21%, and a year ago the best agent managed 2.5%. Obsolete author Garrison Lovely called it by far the most alarming score: still only a leading indicator, but a reminder of how fast change comes under exponential growth.
At max effort it scored 65, five points ahead of Gemini Carbon, six ahead of Opus 5.5 and eight ahead of GPT-6 Astra. The extra-high and high settings both scored 62, filling the next two spots. Gains came mostly from science, knowledge and coding; agentic knowledge work and terminal coding sometimes only reached Opus 5.5 levels.
Anthropic cut Fable's per-token prices by 20%, to $8 input and $40 output per million. But at max effort it costs $26.54 per task across the AA suite: more than double Fable 5.1, 2.6× Opus 5.5 and over 4× GPT-6 Astra. Dropping to high effort brings that to $9.08, 10% cheaper than Opus 5.5 on max.
Anthropic's guidance: medium is the default, high is enough for everyday coding, and extra high is for long-running agentic work with measurable outcomes. Max is for finding the ceiling. Early tester Matt Schumer argues even low covers most tasks, so real-world costs may look better than the headline numbers.
Claude writes 90% of Anthropic's code changes under human supervision. Fable 5.2 is brought in for stubborn bugs, performance work and high-value projects that would otherwise be too slow or costly, such as migrating a critical production service from Python to Rust. That may be good guidance for everyone else too.
Over eight days, the model coordinated more than 1,000 sub-agents to write, test and review an OS. It boots on a laptop and has its own core, drivers, file storage, windows, a login screen, an installer and built-in Claude. Theo's reaction, "Yeah, and?", shows how quickly we're getting desensitized.
Fable 5.2 solved two open theoretical-physics problems about black holes and accurately modeled 28 of 34 gene-therapy capsids. It also fully formalized Fermat's Last Theorem into a machine-readable proof in 30 hours with 1,000 agents; Opus 5.5 stalled 26% of the way through after 18 hours. NLW doesn't think the heavy science emphasis is accidental.
In Anthropic's demo, Fable 5.2 piloted a pair of robotic arms, including on tasks it was never explicitly trained for. It outperformed most models designed specifically to drive robotic arms, which suggests the bitter lesson applies to advanced robotics too.
Fitting Mythos's origin story, testers found the commercial model very strong on security. One found 10+ severe vulnerabilities, created earlier by Opus 5.5, without even asking. On Pavel Hurin's bug hunt bench, Fable found 58 of 105 planted bugs. Most top models find 40–45, and Sonnet 5.5 Max found 51 using twice as many turns at higher cost.
Riley Brown had Fable 5.2 create a 60-second ad for itself live on stream, in one shot, with zoomed-in visuals, artistic shots and realistic lighting and physics. Even Claude Code creator Boris Cherny said "Whoa." NLW sees it as more evidence that agentic video is real.
Every's review, "Fable 5.2 is a paradigm shift if you know what to do with it," describes a teammate building a working video editor that combines Premiere, After Effects and Ableton-style mixing with an AI agent. Their claim: Opus 5.5 keeps you in the loop, but Fable 5.2 "can take the whole product." That reopens the SaaSpocalypse question many had settled in favor of buying over building.
NLW had the same experience with the original Fable 5. Earlier models let him start lots of projects, but Fable actually completed them. He rebuilt the AI Daily Brief website and extraction pipeline in the 48 hours he had access, and barely touched it during the shutdown.
Every's Dan Shipper called it "a spiky kind of nuts." Some tasks will blow your mind like nothing before. For everything else it's sort of normal, and Opus 5.5 is the better choice.
— Ethan Mollick, "Fable 5.2 and Jagged AGI". In "Fable 5.2 and Jagged AGI," Mollick calls the model almost AGI. Its failures have changed: no longer made-up facts but a joyless course, a 182-page paper and creative work buried in rules. Its critics and fact-checkers caught errors but never asked whether the work was too long or no fun.
Asked in the regular Claude interface to show off its capabilities, Fable made things that are hard to make but easy to check. These included a sentence that counts its own letters, music that harmonizes when played backwards, and a hand-built QR code. NLW found them underwhelming, since most people can't tell why they're hard. In Claude Code, it produced a far more interesting three-minute film generated entirely in the browser.
NLW's coding-harness result closely matched what Ethan Mollick got from nearly the same prompt, right down to the twist ending. In both, the first letters of the script's sentences spelled a hidden message. Mollick's spelled his blog name; NLW's spelled the prompt's last line, "Be amazingly clever about this."
Asked for an impressive, data-driven AI Daily Brief episode, Fable proposed timing every AI launch from announcement to availability. It found a real gap in the coverage but couldn't tell that nobody cares. Its backup idea, grading AI predictions that have come due, worked better, but NLW thinks Opus 5.5 would have done roughly as well.
Fable 5.2 asks users to do things that weren't possible before, not just to improve existing work. Its game-building looks like a real step: testers say it handles actual game mechanics, not just good-looking shells. Anthropic's Sholto Douglas called it the first model to pass his bar for "seriously fun and beautiful games." NLW wonders whether games become a new medium for personal and professional work.
Skeptics call Fable models stubborn, heavy on guardrails, short on taste, expensive and token-hungry. Some argue that's why this is 5.2 and not 5.5: no Opus-5-to-5.5-style leap. Within an hour, Theo was already struggling to decide which model to use.
Anthropic's Alex Albert advises: - Ask what's wrong with your idea before it builds anything. - Set approval points up front, such as before deleting things, and make clear the rest is its call. - Start on medium or high effort. - Hand it a whole project and let it run for hours. Anthropic's updated prompting guide adds that when told the codebase is its own, the model merges duplicate code and fixes nearby problems instead of just reporting them.
Even with Opus 5.5, it took three or four weeks for people to start understanding what made it different. NLW suspects Fable 5.2 has something similar in it, but finding it will take experimentation.
