Sign in Subscribe
news

2026-11-10 Claude Fable 5.2 - cheaper per token, twice the cost per task at max effort

Sebastien Dubois

Sebastien Dubois

— 5 min read

Canonical version: 2026-11-10 Claude Fable 5.2 - cheaper per token, twice the cost per task at max effort.

Yesterday (9 November), Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, ten weeks after Claude Fable 5.1. Same structure as before: one model, two levels of safeguards. Fable 5.2 is for everyone, and Mythos 5.2 is for vetted cybersecurity and life sciences organizations.

On paper, it's an easy upgrade. Tokens are 20% cheaper ($8 / $40 per million, down from $10 / $50), output is up to 30% faster, and it tops almost every benchmark Anthropic published. But what you pay also depends on how many tokens a task takes, and at max effort Fable 5.2 uses more than twice as many as Fable 5.1.

In this piece, I'll go over what's new, where the cost goes, and how I'd choose between Fable 5.2 and Claude Opus 5.5.

What's new

The biggest gain is on long agentic work. On the Remote Labor Index, Fable 5.2 scores 41.3%, against about 20% for Opus 5.5, Fable 5.1 and GPT-6 Astra. On FrontierCode 2.0, it reaches 49.1% (Opus 5.5: 45.5%). Gemini 4 Carbon still wins AutomationBench (53.9 vs 51.4).

Anthropic's demos are about scale:

  • One engineer built a prototype operating system from an empty folder, with more than 1,000 subagents over 8 days. It boots on a laptop
  • 1,000 Fable 5.2 agents formalized Fermat's Last Theorem in Lean in 30 hours. The first formalization, by an unreleased research model with a custom harness, took 11 days. Opus 5.5 in the same setup reached 26%, then stalled
  • With Diode, it designed three hardware devices that worked without electrical modifications

The safeguards block about 55% fewer requests than Fable 5.1. Blocked cybersecurity queries are still answered by Claude Opus 4.8, and biology queries by Claude Opus 5.

And Anthropic still calls Opus 5.5 its "strongest daily driver". Fable 5.2 is for the bigger tasks.

Where the money goes

Artificial Analysis ranks Fable 5.2 first on its Intelligence Index (65, at max effort). It also measured $26.54 per index task, about 46 minutes per task (the median is 12.6), and 275M output tokens to run the whole index (the median is 72.8M). They call it "particularly expensive" and "notably slow per task".

On Hacker News, one commenter compared it with Fable 5.1: output tokens per task went from 118k to 274k, so the cost per task doubled despite the price cut. Another pointed out that the charts show max effort by default, a setting "almost never used in practice": "Fable 5.2 on medium looks quite attractive compared to Opus 5.5 high/xhigh. Fable 5.2 on max (the default) looks like a good way to light your money on fire." At high, it scores 62 for about a third of the cost.

Simon Willison hit the other problem with max: his pelican SVG test came back empty, because the model spent its whole 128,000-token budget thinking. Same bug as on Opus 5.5.

Anthropic's own numbers point in the same direction. At high effort, Fable 5.2 matches Opus 5.5 at max on FrontierCode, for about half the cost per task. So the price per token tells you little. The effort level decides the bill.

What people built

Ethan Mollick's review is titled "Fable 5.2 and Jagged AGI". He thinks it's "almost" AGI under a narrow definition, but the gaps moved from errors and hallucinations to taste and judgment, like knowing when its own work is bad. His agents built 450+ 3D worlds and wrote themselves an 87,500-word rule book along the way. And "it did not notice any of these problems itself".

At Every, the Everycut video editor went live in 3 days. Dan Shipper warns that it's "slow and burns through tokens". His advice: give it a big job, run it overnight, and define "done", the budget and how much checking you want.

Lex Toumbourou let it run overnight in Claude Code and got a 12-stage beat 'em up for his nephew, who said: "It's exactly as I imagined." The model wrote its own bot to play every stage and check screenshots.

Kieran Klaassen (Every) had it build "First Crush", a game to learn winemaking: you crush 250 kilos of fruit, read the sugar with a hydrometer, weigh 15 grams of sulfite, pick a yeast, and the game explains why at each step. "The animations still need work. I think it playtested itself for hours." His advice: "Pick a thing you love, make a game about it, and go learn shit."

Launch offers

Pro and standard Team seats get all Fable models for one week (until 16 November), up to 50% of the weekly limit, plus one Fable limit reset. Max, Ultra and premium seats get one to three Fable resets depending on how long they've been subscribed (less than 6 months: 1; 6 to 12 months: 2; more than 12: 3), valid until 8 March 2027. A reset restores only the Fable part of your weekly limit.

One HN commenter summed up the reaction: "I'm going to need a fable reset to explain to me the rules of fable resets and usage."

What I'd do

  • Keep Opus 5.5 at medium as the default. That's also what Anthropic's developer guide recommends
  • When Opus 5.5 falls short, follow the guide's order: first give the model a way to verify its own work, then raise effort by one level, then move to Fable 5.2 at a lower effort. Don't jump to max
  • Use Fable 5.2 for long unattended runs (migrations, big refactors, research) with a written definition of "done"
  • If you let it refactor, add the ownership sentence from Anthropic's prompting guide ("This codebase is yours..."). Without it, the model adds code next to the existing structure instead of cleaning it up. With it, expect about 1.5x the output tokens and some unrequested changes, so review the diff, not only the summary
  • Reread the rules and memory files your agents carry. Fable 5.2 follows cautious instructions closely, and one stale "never retry" note was enough to stop it from retrying in more than half the tries
  • On Max or Ultra, keep your resets for a big job. They're valid until March

The full breakdown (benchmarks, safeguards, prompting tips, launch offer rules) is in my Claude Fable 5.2 note.

That's it for today! ✨

References

About Sébastien

Ready to get to the next level?

Found this valuable? Share it with someone who needs it.

Join 6,000+ readers. Get practical systems for knowledge & AI. Free.

Subscribe ✨

Free: Knowledge System Checklist

A clear roadmap to building your own knowledge system. Subscribe and get it straight to your inbox.

6,000+ readers. No spam. Unsubscribe anytime.

Subscribe