OpenAI cuts GPT-5.6 Luna 80% after a model rewrote its own inference code

OpenAI cut the price of GPT-5.6 Luna by 80 percent, to $0.20 per million input tokens and $1.20 per million output tokens, and dropped the larger Terra tier by 20 percent. The discount is not the interesting part. What is interesting is how they got there. OpenAI says another model in the family, GPT-5.6 Sol, rewrote its own production inference kernels using Codex, working directly in the Triton and Gluon GPU languages. Sol looked for work that could be precomputed, skipped, or run in parallel, and cut memory movement and synchronization stalls. Together that took about 20 percent off the end-to-end cost of serving the models, which OpenAI passed through to the price.

At the new rate, Luna is roughly five times cheaper on input than Anthropic's Claude Haiku 4.5 and undercuts Google's Gemini 3.1 Flash-Lite, as Simon Willison notes in his write-up. The mechanism is the story worth watching: a model handed the job of making itself cheaper to run, with the savings landing on the invoice rather than in a research paper. Kernel optimization is exactly the kind of narrow, well-specified, heavily testable task where an agent can grind without much room to go wrong.

Why it matters

If you run high-volume workloads on a cheap model tier, Luna's new price changes what you can afford to send to an API instead of a local model. It is worth re-pricing your current provider against it before your next billing cycle.