AGENTS.md sent Codex to Notion on every image, costing 6万トークン

The developer was generating images for short videos by calling Codex CLI from Claude Code. Because Codex reads AGENTS.md every time it runs, an instruction written for conversations—to read a Notion profile before answering—was also executed during image work. The work log put the cost at about 60,000 tokens per image, though the developer notes this figure was not re-measured from logs and should be treated as an estimate.

The model comparison tested terra (gpt-6.1-terra) and luna (gpt-6-luna) at low reasoning, using the same prompts and three reference images. The weekly usage gauge moved only from 39% to 40% after ten images, so the developer compared total tokens in Codex logs instead. In the five-cut comparison, terra stayed steady at around 90,000 tokens per image, while luna ranged from about 109,000 to 192,000. The developer suspects luna called the image tool two to four times per image, resending prior context as input each time, but notes this is inference and the call counts per cut were not recorded.

For reference, an earlier run with the higher-tier astra model used about 0.35–0.37% of the weekly quota per image, though that was before the Notion reads were stopped, so the difference is not purely model-related. The developer also notes that one day images took 7–18 minutes, which appeared to be OpenAI-side queueing rather than the model.