# Claude Fable 5.2 Sparks AGI Talk as Builders Test Its Limits

> Anthropic’s latest model is generating ambitious software, research and 3D demos. Early reactions on X show excitement about its independence—and disagreement over how reliably and affordably it works.

By Lidia Yadlos · November 10, 2026

Canonical: https://blockster.com/claude-fable-5-2-agi-builders-opus-costs

Claude Fable 5.2 has been public for barely a day, and the conversation has already moved beyond whether it writes better code. Early users are asking how much of an entire project they can hand over—and whether its growing range of abilities deserves the label artificial general intelligence.

[Anthropic released Fable 5.2 and Mythos 5.2 on November 9](https://www.anthropic.com/claude-fable-and-mythos-5-2). The company reports progress across coding, scientific research and engineering, including working electronic-device designs and solutions to two open problems in black-hole physics.

The strongest AGI argument comes from Wharton professor [Ethan Mollick](https://www.oneusefulthing.org/p/fable-52-and-jagged-agi), who tested Fable before launch. His assessment is “almost” AGI: impressive across many intellectual tasks performed on a computer, with important gaps remaining.

That qualification matters. The original posts reviewed by Blockster show enthusiastic demonstrations alongside complaints about prompting, costs and usage limits. They do not establish a community consensus that AGI has arrived.

## What Has Actually Changed?

Anthropic says the new models generate output up to **30% faster** than their predecessors. It also reports that approximately 1,000 Fable agents formalized Fermat’s Last Theorem in 30 hours—translating an existing mathematical proof into a form a computer can check.

Fable and Mythos share the same underlying model. Fable includes stronger restrictions around cybersecurity and biology; Mythos access is available through verification programs for organizations working in those fields.

The [model documentation](https://platform.claude.com/docs/en/models/fable-5-2/overview) describes Fable as suited to work that can run for hours, from research to finished documents and software projects. It supports a one-million-token context window, allowing large amounts of material to be supplied within a session.

Its standard API prices also show where it sits in Claude’s lineup:

| Model | Input per million tokens | Output per million tokens |
| --- | --- | --- |
| Claude Fable 5.2 | $8 | $40 |
| Claude Opus 5.5 | $4 | $20 |
| Claude Sonnet 5.5 | $2 | $10 |

These are standard input and output rates, before caching, batch discounts or other applicable charges. They are developer API prices, not monthly subscription prices.

## On X, the Excitement Is About Finishing the Job

AI reviewer **Matthew Berman** described Fable 5.2 as the best model he had used after extensive early testing. In his November 9 post, he praised its ability to make useful decisions with limited direction, find bugs and build 3D experiences.

His examples included a game inspired by Gears of War and work in Unreal Engine and Blender. These are Berman’s reported demonstrations, rather than independently audited evidence that Fable can reproduce every feature of a commercial game.

His most revealing observation was practical: he said he rarely needed follow-up prompts to persuade it to finish.

[Post from @MatthewBerman on X](https://twitter.com/MatthewBerman/status/2119871207206158459)

That is a meaningful distinction for people building with AI. A system that produces a promising first attempt still leaves someone managing the unfinished pieces. Completing more of the work could reduce that burden—provided the result survives review.

Independent evaluations give some support to the capability gains. [ARC Prize’s verified results](https://arcprize.org/results/anthropic-claude-fable-5-2) show Fable reaching **81.2% on ARC-AGI-3 with the standard testing setup**, and **97.7% with a provider adapter that preserves more of its reasoning state between requests**.

Those scores come from different configurations. They should not be treated as interchangeable, or as a certificate that general intelligence has been achieved.

## The Human Judgment Gap

Mollick tested work spanning business planning, hardware and research. In one experiment, Fable developed a proposed service that would alert equipment salespeople to municipal projects before they reached the bidding stage. He found the idea credible, but stopped short of letting the AI launch and operate it.

Another project exposed a stranger weakness. Asked to improve a collection of generated 3D worlds, the agents created increasingly elaborate approval procedures. The rulebook eventually reached **87,500 words**, slowing the work.

The problem was recognizing when its own process had become counterproductive. Mollick’s definition also excludes much physical and interpersonal work, making “almost AGI” a narrower claim than an AI capable of doing most jobs.

> “Most critically, the AI did not notice these problems itself.”
>
> — [Ethan Mollick, Wharton professor, in ](https://www.oneusefulthing.org/p/fable-52-and-jagged-agi)_[Fable 5.2 and Jagged AGI](https://www.oneusefulthing.org/p/fable-52-and-jagged-agi)_

That leaves a substantial role for the person setting priorities, judging quality and deciding when the work is finished. More output does not automatically mean a better result.

## Better Results Can Still Come With a Bigger Bill

Developer **Tony, posting as** **[@EnvolDev](https://x.com/EnvolDev/status/2119907318729245180)**, shared a useful comparison from a single creative task: an ice-cathedral scene built using the same prompt.

He reported Fable completing it in roughly **1 hour 57 minutes for $81.10**, compared with approximately **3 hours 11 minutes for $45.97** using Opus 5.5. He preferred the Fable result and highlighted its shorter completion time.

That is one user’s comparison, not a general performance guarantee. It nevertheless illustrates the calculation builders face: a higher bill can be worthwhile if the output saves enough time or requires less correction.

Even Berman’s enthusiastic review ended with a qualification. He expected to keep using Opus for much of his everyday work because of subscription quotas, switching to Fable for planning the hardest jobs.

Other users reported more friction. In a November 9 post, **Ben Kaufman** said his first few hours with Fable required longer, more specific prompts and that he found its understanding of intent and initiative disappointing.

[Post from @ben\_kauf on X](https://twitter.com/ben_kauf/status/2119927997058924723)

These experiences can coexist. A model may handle an ambitious graphics project well and frustrate someone trying to make a precise change in an existing workflow. Early posts are useful examples of that variation; they are not a representative survey.

## Use Opus for the Build, Fable for the Hardest Finish

You do not have to run an entire project through Fable to benefit from it. [Anthropic recommends starting with Opus](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2) for most workloads, then moving to Fable when Opus at higher effort still falls short.

A practical way to apply that advice is to let **Opus 5.5 handle the first draft, routine coding and revisions**, then give **Fable 5.2 a focused handoff**: the current work, the requirements, what has already been checked and the remaining problems. Fable can then tackle a difficult bug, review the implementation or complete the parts Opus could not finish satisfactorily. That is an application of Anthropic’s broader guidance, not a requirement to switch models at the end of every task.

The potential saving is in **cost and premium-model usage**, not necessarily fewer tokens overall. Opus’s standard API input and output rates are half Fable’s. Keeping routine iterations in Opus can therefore reduce spending compared with doing everything in Fable, but rereading context, repeating work and lengthy reasoning can eat into that saving. Subscription usage limits are separate from API billing.

## What This Means for Builders and Businesses

The commercial question is how much dependable work a company gets for its money.

For a founder, that could mean reaching a working prototype sooner. For an engineering team, it could mean completing a difficult migration with fewer rounds of supervision. In both cases, the comparison should include the cost of checking and correcting the result.

The release also connects with a question raised in [Blockster’s report from Google DeepMind’s Miami AI Build Night](https://blockster.com/miami-ai-build-night-google-gemini-claude): where does each model actually help people do better work?

And as companies give agents greater responsibility, checking the information behind their actions becomes more valuable. That is the problem explored in [Blockster’s coverage of Space and Time’s private, verifiable data platform](https://blockster.com/space-and-time-enterprise-private-data-verifiable-ai). Verifying a calculation and judging an AI’s overall decision remain different tasks.

Fable 5.2 makes the AGI debate more tangible because people can point to projects it helped complete. The next test is whether those results hold up across ordinary users, real budgets and work that has to be right after the demo ends.