Overview
- GPT‑5.6 Luna is now widely available at dramatically lower token rates, with OpenRouter listing about $0.20 per million input tokens and $1.20 per million output tokens which fall much further on cache hits and batch tiers.
- OpenAI announced a major price shift that cut Luna roughly 80% from the prior tier, a move first signaled by Sam Altman and reflected in public API and gateway pricing that changed unit economics for many workloads.
- The cost fall comes from architectural and inference changes: mixture‑of‑experts models activate a small subset of a large parameter pool to lower runtime compute, and prompt caching plus batch APIs create tiered pricing where cache hits can cost an order of magnitude less.
- Those lower marginal costs make previously uneconomical designs practical, including ten‑step agent loops, naive retry and verification cycles, and consumer features such as personalized daily news that move from dollar‑scale to cent‑scale per run.
- Important limits remain: higher‑cost Sol‑class frontier models still outperform Luna on deep reasoning and some security tasks as shown in Semgrep tests, and adopting the cheap stack requires engineering work on caching, routing cold starts, validation, and safety controls which will shape who shifts traffic to lower‑cost models.