OpenAI cuts GPT-5.6 Luna's price by 80 percent

OpenAI has cut the price of its cheapest model sharply. In its price-performance post, the company says GPT-5.6 Luna now costs $0.20 per million input tokens and $1.20 per million output, an 80 percent reduction, while the larger GPT-5.6 Terra gets a smaller 20 percent cut. The new Luna price sits under Google's Gemini 3.1 Flash-Lite, at $0.25 and $1.50, and leaves Anthropic's Claude Haiku 4.5, at $1 and $5, about five times more expensive on input.

The more interesting part is how OpenAI got there. The company attributes the savings to GPT-5.6 Sol, a model that it says autonomously rewrote and optimized its own production inference kernels, improving GPU memory use and parallelization to take roughly 20 percent off end-to-end serving costs. So the price cut is partly a story about a model tuning the stack that serves it.

Simon Willison, who follows model pricing closely, treated the drop as a real shift for low-cost models and moved his agent.datasette.io demo from Gemini to Luna on the strength of it.

Why it matters

If you run high-volume, latency-tolerant work on a budget model, Luna is now the cheapest credible option in its tier, and it is worth benchmarking against your current Gemini Flash-Lite or Haiku setup before the next billing cycle.