Short RL phase on GLM 5.3 Flash costs $372 vs $1,894 for notes, breaks even in 8.7 days

teortaxesTex · x · 2026-11-05

EngramLab shared a real cost experiment: a short RL phase fine-tuned GLM 5.3 Flash to actively use its notes instead of stuffing them into context.

Key numbers:

Generating the notes cost $1,894.49 in inference, while the follow-up RL training cost only $372.57;
Per the cumulative cost curve, the training route breaks even with the notes-in-context route after about 8.7 days / 3,487 queries, then stays 5-10x cheaper.

The takeaway: with good infrastructure, replacing long-context inference overhead with training is surprisingly cost-effective — "training is just inference with a backwards pass." A concrete, replicable cost accounting for anyone designing agent memory or knowledge workflows.

Original post →

Short RL phase on GLM 5.3 Flash costs $372 vs $1,894 for notes, breaks even in 8.7 days

More from coding & agent