OpenAI Cuts Luna Price 80%, Credits AI-Rewritten Kernels

OpenAI cut GPT-5.6 Luna's price 80% on July 30, to $0.20 per million input tokens and $1.20 output.

That undercuts Gemini 3.1 Flash-Lite's $0.25 input rate and puts Luna at a fifth of what Claude Haiku 4.5 charges. Anyone routing between models on cost just watched the cheapest option change again.

Simon Willison moved his agent.datasette.io demo off Flash-Lite to Luna the same day. He then made Luna the default model in LLM 0.32rc2, replacing GPT-4o mini.

OpenAI says GPT-5.6 Sol rewrote its own production kernels in Triton and Gluon, unsupervised, and claims that cut serving cost 20%.

I've heard the pitch that models will eventually improve the infrastructure running models. This is the first time I've seen it land on a price sheet.

Twenty percent off serving cost doesn't explain an 80% price cut alone. Most of that gap is OpenAI matching Gemini Flash-Lite and the flood of cheap open-weights models.

But kernel optimization is narrow, benchmark-checkable work, exactly where an agent with a fast feedback loop should beat a person on throughput. Nobody's disputed the specifics yet.

Artificial Analysis scores Sol at 59 on its Intelligence Index, one point under Claude Fable 5, at roughly a third the cost per task.

A day later, DeepSeek shipped V4-Flash-0731: MIT-licensed weights, 304B parameters, $0.06 per million tokens, scoring 50 on the same index, a point behind Luna. It's specifically adapted for Codex, close to a drop-in swap inside an OpenAI-shaped harness.

Vercel's AI Gateway passed the July 30 cut through at zero markup, so anyone routing through it pays less with no code change. Static routing logic tuned to the old price curve is now picking the wrong models regardless.

If you picked models by price three months ago, that decision is stale now.