Claude Fable 5.2, one day later [video]

Henrik Kvamme on X: "Fable 5.2 tops the benchmarks. So I had it make its own honest launch video, from 160 day-one user reports. How to use it well:

• Start with Opus 5.5
• Give Fable the big, vague job
• Set effort to high, not max
• Don't let it grade its own work Why, in the video:" / X

Henrik Kvamme on X: "Fable 5.2 tops the benchmarks. So I had it make its own honest launch video, from 160 day-one user reports. How to use it well:

• Start with Opus 5.5
• Give Fable the big, vague job
• Set effort to high, not max
• Don't let it grade its own work Why, in the video:"

Fable 5.2 tops the benchmarks. So I had it make its own honest launch video, from 160 day-one user reports. How to use it well:

• Start with Opus 5.5
• Give Fable the big, vague job
• Set effort to high, not max
• Don't let it grade its own work Why, in the video:

Fable 5.2 tops the benchmarks. So I had it make its own honest launch video, from 160 day-one user reports. How to use it well:

• Start with Opus 5.5
• Give Fable the big, vague job
• Set effort to high, not max
• Don't let it grade its own work Why, in the video:

3/ The catch: at API prices, those builds cost $11,000 to $15,000 each. And the usage in a $200 plan is worth about one fifth as much on Fable 5.2 as on Opus 5.5.

4/ On 30 ordinary coding tasks, two private tests found a tie with Opus 5.5. Fable cost almost twice as much per solved task. The gap shows only on big, hard jobs.

5/ Sources: game footage and cost estimates @MatthewBerman, plan math @theo, bug-hunt test @PawelHuryn, paper example @emollick, effort data @ArtificialAnlys. Not made or endorsed by Anthropic. Research: six Sonnet 5.5 agents. Script and animation: Fable 5.2.

Using real user reports as the script is a great idea, way more honest than the usual hype reel. As a motion designer making launch videos, I'm curious: how much did you art-direct the visuals vs. let it run?