Claude Fable 5.2 Just Took First Place at 64.6, Costs 20 Percent Less Than Its Predecessor, and Spent Eight Days Building an Operating System With 1,000 Agents


One model, two doors

Anthropic released two models on November 9, and the most important fact about them is that they are the same model. Claude Fable 5.2 is open to everyone on paid Claude plans, the Claude Platform API, AWS, Google Cloud and Microsoft Azure. Claude Mythos 5.2 is the same underlying system without the restrictions on sensitive work, and it is available only to organizations that pass Anthropic’s Cyber Verification Program or Life Sciences Verification Program.

Fable carries safeguards that block potentially dangerous cybersecurity and biology requests. Mythos keeps those capabilities for verified teams doing approved security or biomedical research. Access policy is no longer a wrapper around the product. In this release, access policy is the product split itself.

The API identifier for the public model is claude-fable-5-2. The context window stays at 1 million tokens and the maximum output stays at 128,000 tokens, both unchanged from Fable 5.1, which remains available.

First place, and the bill that comes with it

Independent measurement gives the launch its headline. On the Artificial Analysis Intelligence Index v5, in the November 9 edition, Fable 5.2 at its maximum setting scores 64.6 and ranks first. That is 4.6 points above the previous leader, Google’s Gemini 4 Carbon at 60.0, and 14.6 points above Fable 5.1’s best score of 50.0. Even the high setting, at 61.8, beats Carbon.

Price moved the other way from capability. Standard pricing is $8 per million input tokens and $40 per million output tokens, with cached input reads at $0.20, down 20 percent from Fable 5.1. Anthropic says output is up to 30 percent faster than the previous version, and its model documentation now lists latency as Moderate instead of Slower. An optional fast mode, which can raise output speed by up to 2.5 times, costs double: $16 input and $80 output.

The catch sits in cost per finished task, the number Artificial Analysis calculates from how many tokens a model actually uses to complete its test work. At maximum effort, Fable 5.2 costs $26.54 per task, the highest figure in the table. At the high setting it costs $9.08 per task, about 10 percent less than Claude Opus 5.5 at its maximum. The lesson matches the one enterprise buyers keep relearning this year: a cheaper token price does not guarantee a cheaper job, because a stronger model that thinks longer can spend the saving and more.

The company’s own scoreboard, read carefully

Anthropic’s benchmark table puts Fable 5.2 first among five models on 13 of 15 metrics. On FrontierCode 2.0, a hard coding test, it scores 49.1 percent at the extra high setting, up from 32.4 percent for Fable 5.1 and above Opus 5.5 at 45.5 percent. Two rows where it does not lead are worth naming: Gemini 4 Carbon leads on the AutomationBench workflow test at 53.9 percent, and Opus 5.5 leads on the AA-Briefcase long workflow ranking at 1,800 Elo. The two cybersecurity rows and the protein design row in the table use Mythos 5.2 scores, not Fable scores, because Fable is the version that restricts exactly that work. All of these figures are self reported by Anthropic.

One safeguard number cuts the other way from what readers might expect. Anthropic says Fable 5.2’s safeguards block about 55 percent fewer requests than Fable 5.1’s did. Read plainly, the new public model refuses less often while scoring higher. Whether that balance holds in real use is a question only independent testing over the coming weeks can answer.

Anthropic also undercuts its own flagship in its guidance. The company still recommends starting most everyday coding and knowledge work on Opus 5.5, and reaching for Fable 5.2 on the hardest, longest technical jobs, or using it to direct smaller models. That is an unusual thing for a launch post to say, and it is probably the most useful buying advice in it.

An operating system, two black hole problems, three devices

The demonstrations are the part of the launch designed to travel, and they should be read as company reported results, not independently verified ones.

Anthropic says that during testing, an engineer asked Fable 5.2 to build an operating system from scratch. The model coordinated more than 1,000 AI agents over eight days and produced a working system that could boot on a laptop, with windows, file storage and a login screen. The company also reports that the new models helped solve two previously open problems in black hole physics, and designed three working electronic devices.

Each claim points at the same bet: frontier value is moving from single answers to long, managed assignments that run for hours or days, with one model planning and many agents executing. An eight day build is a glimpse of that direction. It does not establish that the model can independently deliver an arbitrary real world engineering project, with real users, real deadlines and real consequences for failure. No outside party has yet repeated the operating system run or graded the physics results in public.

The timing gives the demos extra weight. Five days earlier, the independent evaluator METR reported that Anthropic’s rate of AI progress has accelerated about 1.9 times since October 2025, with AI itself doing much of the speeding up. A model that coordinates a thousand agents to build software is exactly the kind of system that assessment was describing. The launch also lands in the week Anthropic is expected to set its IPO price range, according to the Wall Street Journal, and it became the top story on Hacker News on launch day. Capability, safety framing and market timing are all in the same announcement, and readers should keep them separate even when a launch does not.

What to do with this, this week

  • Test at the high setting first. At $9.08 per Artificial Analysis task, high is the value tier. Reserve maximum effort for jobs where a failed run costs more than the tokens.
  • Budget per finished task, not per token. Separate input, output, cached context and fast mode traffic in your own logs before you migrate any workload. Fast mode doubles the token price.
  • Keep Opus 5.5 as the default. Anthropic itself recommends it for most coding and knowledge work. Move only the hardest, longest jobs to Fable until your own results say otherwise.
  • Do not plan security or biology work around Fable. Those requests are restricted by design. Teams that need that capability need Mythos verification first, and approval should come before any workflow is built on it.
  • Treat the demos as directions, not guarantees. Run your own long horizon test, with tool failures and repeated attempts included, before trusting an eight day agent run with anything that matters.

The bottom line

Claude Fable 5.2 is a real jump on the public record: first place at 64.6 on an independent index, 14.6 points above its predecessor, 20 percent cheaper per token, and up to 30 percent faster. It is also the clearest example yet of where frontier releases are heading. One model now ships through two doors, a public door with safeguards that refuse less than last time, and a locked door for verified specialists. The score is no longer the whole product. Who gets access, at what effort setting, at what cost per finished task, and with what proof behind the demos, is the product. Fable 5.2 leads the table today. The independent, repeated run testing over the next few weeks will decide how much of that lead survives contact with other people’s work.