Astra vs Claude on a design task: the 3D lightsaber

This is the lightsaber reel. I wrote one brief for an interactive 3D lightsaber in a single HTML file: built from basic shapes, a glowing blade, parts that slide out when you click them, an explode-all button and a colour picker. I gave the same brief to Claude Fable 5.1 in Claude Code and to GPT-6 Astra through my Hermes agent.

This one was not a single attempt. Each side got art direction from me and went through three versions. What you see in the reel is the third version from each, untouched.

The reel shows Fable at $37.61. That figure was wrong: my first pass counted some responses two or three times. Counted properly, Fable came to $19.29 across the three versions. Astra came to $6.50 on an identical price per token, on far fewer tokens: about 3.5 million cached tokens read against 9.5 million, and about 29,000 output tokens against 101,000.

It is still not a clean comparison. Astra's log does not keep its reasoning tokens, my Claude Code setup adds a lot of context to every call, and Fable checked its own work in a browser 45 times. My best reading is that Astra was two to three times cheaper for this job, and clearly the more token-efficient of the two.