GPT-6.2 Astra as planner only: same 3D scene for $1.62 instead of $7.40

GapNew4766 · reddit · 2026-11-18

A developer tested a cost-saving pattern for long agent runs: keep GPT-6.2 Astra strictly as the planner and delegate coding and frame checks to a locally-run Qwen 3.8 27B, with Astra only receiving short summaries back.

Setup: one hard prompt run twice — a 25-second "birth of a sun" Three.js scene in a single HTML file, with the model required to open the page and verify 12 key frames.

Why the bill is big: Astra costs $8/1M input and $40/1M output on OpenRouter. In an agent loop, almost all cost is input — every step rereads the task, thousands of lines of code, and frame-check screenshots. Getting summaries instead cut the bill 4.6×.

Honest tradeoffs: 3× slower (local tokens); scenes are not identical — judge the video yourself; it's one task, not a benchmark. The author keeps this setup for anything that can run overnight: Astra's judgment is worth paying for, its typing isn't.

The author works on the open-source agent Atomic Agent; the mode is called Fusion, and the full prompt is available on request.

Original post →

More from coding & agent

coding & agent channel →