Anthropic lässt die Bombe platzen: Fable 5.2 in 5 Praxistests

Summary

The video compares Anthropic’s newly announced Fable 5.2 with Opus 5.5 through five hands-on coding tasks. The presenter focuses not only on benchmark claims, but also on whether the models’ differences matter in everyday development.

Anthropic’s claims and model context

The presenter reports that Anthropic introduced both Fable 5.2 and the more restricted Mythos 5.2. Mythos is described as intended for selected organizations, while Fable is available for general use.

The video cites strong benchmark results and Anthropic’s claims that its models are used extensively for software development and can assist with difficult engineering and research problems. These are presented as company claims, not as results independently tested in the video.

How the comparison was run

Both models received the same prompts, tools, effort setting, empty project folders, and 25-minute limit per task in Claude Code. The presenter disabled plugins and custom instructions, tested each result, and then gave both models the same correction instruction: find and fix their two worst mistakes. The results were initially shown anonymously as A and B.

The five tasks were:

1. Build a functional cybersecurity landing page.
2. Create a 25-second Remotion animation.
3. Make a playable game.
4. Build a gravity and three-body physics simulation.
5. Create a browser-based 3D vehicle configurator.

Results

• Landing page: Both versions looked polished and passed the presenter’s functional checks, including mobile layout, scanning interactions, pricing toggles, and FAQ. The result was effectively a draw. A started coding quickly and tested more; B spent longer planning and produced substantially more code.
• Remotion animation: Both produced compelling videos. A communicated the agents’ activities more clearly, while B was more cinematic. Each had a small visual flaw, and both corrected their issue during the repair round. The scores were very close.
• Game: Both delivered complete, playable games with movement, combat, collectibles, upgrades, win and loss states, and visual effects. B’s version had a stronger atmosphere and a more dramatic lockdown sequence, but was harder to beat. The presenter also noticed that the models’ internal gameplay checks took shortcuts rather than playing through the whole game.
• Physics simulation: This was the clearest difference. The presenter independently checked energy conservation using the simulated bodies’ masses, positions, and velocities. B’s simulation showed much smaller energy drift and more stable orbital behavior. A’s three-body scenario also failed to remain meaningfully chaotic until it identified and corrected the issue during the repair round.
• 3D configurator: Both models built the vehicle from code rather than using premade 3D assets. Their options changed the vehicle’s visible features as well as its price, and the specified base and fully equipped prices worked. B made a more technically detailed vehicle; A presented its options more effectively.

Overall assessment

The reveal is that A was Opus 5.5 and B was Fable 5.2. Fable generally produced more detail and performed better on the physics task, with a slight edge in several other comparisons. However, it often planned longer, generated much more code, and sometimes risked not finishing within the time limit. Opus tended to start sooner, test more often, and make more targeted repairs.

After the correction rounds, the reported API list-price totals were about $16 for Opus and $41 for Fable. The presenter concludes that Opus remains the better-value everyday workhorse, while Fable may be worth its higher cost when precision, detail, or polish is especially important. The presenter emphasizes that this was an informal comparison of five tasks—not a scientific benchmark—and that results may differ on research or hard-logic work.

Main speaker and sources

• Main speaker: The video’s presenter, who designed and evaluated the five practical tests.
• Other sources discussed: Anthropic’s announcements and benchmark claims, along with the presenter’s own functional, visual, and physics checks.