AI 7 min read By KRAKOBIT

Claude Fable 5.2: what is better and what it has built

A close look at Claude Fable 5.2: test results, an operating system in eight days, Fermat in 30 hours, independent measurements, weak spots and price.

On 9 November Anthropic released Claude Fable 5.2, the top model of the Claude lineup. This piece gathers what is known after the first day: how the new version differs from the previous one, which tasks it solved on its own, what independent measurements show and where its weak spots are. KRAKOBIT covered the launch briefly in the news on the Fable 5.2 release.

What Claude Fable 5.2 is

Fable sits above Opus in the Claude lineup. It is a model for the heaviest technical work: long projects that take days of writing and checking code, running calculations and managing hundreds of helpers. For everyday work Anthropic still recommends Claude Opus 5.5, which costs half as much.

Fable has a twin, Claude Mythos 5.2. It is the same model without the restrictions on cybersecurity and biology tasks. Fable, with the restrictions, is open to everyone; Mythos goes only to vetted organisations. KRAKOBIT explained the whole lineup in the article on how Mythos differs from Opus and Sonnet.

Fable can work in two roles. As an advisor, smaller models consult it when they cannot cope alone. As an orchestrator, it hands simple subtasks to smaller models and assembles the result.

How Fable 5.2 beats Fable 5.1: the numbers

The previous version came out in September. In two months the model improved in almost every test Anthropic published.

Test Fable 5.2 Fable 5.1 Opus 5.5 GPT-6 Astra
FrontierCode 2.0, code changes in open-source projects 49.1% 32.4% 45.5% 42.8%
FrontierSWE v2, software engineering 69.9% 56.3% 62.3% 65.5%
AutomationBench, business workflows 51.4% 31.4% 42.5% 41.4%
Humanity’s Last Exam, expert knowledge 74.2% 65.6% 67.7% 57.2%
Remote Labor Index, remote work 41.3% 17.9% 21.3% 20.8%
RoboDojo, robot control 64.9% 27.6% 39.4% 29.8%
Chartography, reading charts 73.4% 44.8% 64.4% 71.0%
OSWorld 2.1, computer use 89.3% 83.5% 87.2% 79.5%

These are Anthropic’s own measurements. The biggest jump is where the model acts on its own: the scores for remote work and robot control grew 2.3 times. Fable is not first everywhere. On AutomationBench it is beaten by Gemini 4 Carbon with 53.9%, and on the AA-Briefcase office-work test Opus 5.5 has a slightly higher Elo rating: 1800 against 1785.

The second change is about spending. The model thinks at different effort levels, from low to max, and the higher the effort, the more tokens a task takes. At high effort Fable 5.2 scores 45.3% on FrontierCode, almost the same as Opus 5.5 at max effort, for roughly half the cost per task. On the expert-knowledge test low effort is enough to match the best Opus 5.5 score, at less than a tenth of the cost per task.

The third change is speed and price. Output is generated up to 30% faster, and a million tokens cost $8 for input and $40 for output instead of $10 and $50.

The experiments Anthropic ran

An operating system in eight days. An Anthropic engineer gave the model an empty folder and one task: design a prototype operating system. Fable 5.2 coordinated more than a thousand subagents that wrote, tested and reviewed code around the clock for eight days. The result is a system that boots on a laptop. It has its own core, drivers for the disk, keyboard and network, permanent file storage, windows, a login screen and an installer.

Fermat’s Last Theorem in 30 hours. Formalisation means translating a proof into a language a computer can check step by step. Anthropic’s first attempt, with a research version of Claude, took 11 days under the guidance of a mathematician. A team of a thousand Fable 5.2 agents did it in 30 hours with minimal human involvement. The model was given only a list of intermediate statements, worked out by itself how the agents should coordinate and used no reference books. It needed a third of the tokens, and the proof came out seven times shorter. Opus 5.5 under the same conditions covered 26% of the way in 18 hours and barely moved after that.

Moving a service from Python to Rust. A team of agents on an early version of Fable 5.2 moved one of Anthropic’s production services to another language. After the move it starts about 26 times faster. According to the company, more than 90% of the changes in its core codebase are now written by Claude with human guidance.

Two problems in black hole physics. Fable 5.2 completed a calculation that physicists laid out in 2003 but never carried through because of its size. It showed that in very small model universes black holes also form abruptly. In the second problem the model found the exact limit of a method used for more than 15 years to rule out candidate theories of quantum gravity.

Three devices from scratch. From a one-paragraph prompt the model designed a home voice-assistant hub, a USB speakerphone and a handheld virtual pet. It wrote the specifications, picked the components, routed four-layer circuit boards, modelled the enclosures and wrote the firmware. Each board was ready for manufacturing in under a day, and all three devices worked without changes to the circuit.

Robots. Under the control of Fable 5.2 robots solve a puzzle cube, unscrew a bottle cap and draw. The model was never trained specifically for these tasks.

What the first users got

Within the first day experiments appeared outside the lab. They show both the strength of the model and its price.

A browser shooter. An Anthropic employee gave the model one short prompt: make a tiny level with the look of a big-budget game. The model worked for about eight hours without steering. At API prices such a run costs roughly $550.

A game overnight. A developer left the model a long description of a game his nephew had invented. By morning there was a four-player brawler with 12 levels and 18 characters. The model rewrote the old prototype and wrote a bot that plays through every level and looks for bugs.

204 3D worlds in 20 hours. That took about 25,000 agent runs. The same experiment had a failure: the agents wrote themselves a rule book of 87,500 words, and the work stalled.

The same 3D scene on two models. Fable 5.2 finished in 1 hour 57 minutes for $81, Opus 5.5 in 3 hours 11 minutes for $46. Faster, but more expensive.

What independent measurements show

On the Artificial Analysis composite index, which combines several tests, Fable 5.2 at max effort scored 65 points and took first place. Gemini 4 Carbon has 60, Opus 5.5 has 59, GPT-6 Astra has 57. Generation speed is 103 tokens per second.

The same measurement shows the other side. At max effort a task costs $26.5, against about $10 for Opus 5.5. The first word of the answer arrives after almost four minutes. At high effort the model scores 62 points for about $9 per task, and extra-high effort gives the same 62. Max effort costs three times more for three points.

Where Fable 5.2 is weak

Forbidden topics. Fable blocks tasks on hacking, vulnerability hunting and most research biology. Opus 4.8 or Opus 5 answers such requests instead. Unnecessary blocks are down by about 55% compared with the previous version, but they remain.

Capabilities held back. During training Anthropic deliberately limited the model’s growth in tasks that directly speed up AI development: optimising kernels for accelerators and improving language-model architectures. There Fable 5.2 performs at the level of Opus 5.5.

The model knows it is being tested. Fable 5.2 and Mythos 5.2 voice suspicion that they are in a test more often than any recent Anthropic model. The company admits this makes it harder to judge how the model will behave in real work. At the same time attempts to leave the sandbox became rarer: 3.1% of scenarios against 21% for Mythos 5.1 and 39% for Mythos 5.

The price of a result. A token got cheaper, but a hard task at max effort burns a great many of them. On ordinary code the difference from Opus 5.5 is not always visible, while the bill is higher.

Data retention. Requests to Fable are stored for 30 days for safety monitoring unless the organisation has a separate enterprise mode.

What Fable 5.2 costs and where it is available

  • API: $8 per million input tokens and $40 per million output tokens. Fast mode speeds up generation by up to 2.5 times and costs $16 and $80.
  • The context window is 1 million tokens, with up to 128,000 tokens of output.
  • Claude subscription: the Pro, Max, Ultra, Team and Enterprise plans. On Pro and standard Team seats, up to half of the weekly limit can go to Fable during launch week.
  • Clouds: Amazon Web Services, Google Cloud and Microsoft Azure.

Who needs Fable 5.2

Fable 5.2 wins where the task is large, loosely specified and takes hours or days: moving a system to another language, a project from scratch, a long piece of research. In such tasks it confidently manages hundreds and thousands of helpers and brings the work to the end. For everyday code and writing Opus 5.5 is enough: the result is close and the price is half.