Advancing the price-performance frontier with GPT‑5.6
Simon Willison's Weblog Simon Willison ● Covered by 14 sources
OpenAI just slashed prices on two GPT-5.6 models, with the cheap one dropping a jaw-dropping 80%. It's now the cheapest frontier-grade model on the market, beating Google and Anthropic on cost.
Based on reporting by Simon Willison's Weblog, Simon Willison — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI dropped a pricing bombshell today, and it's the kind of move that reshuffles the entire cheap-model tier overnight. GPT-5.6 Terra gets a modest 20% cut, but GPT-5.6 Luna is where things get wild: an 80% reduction that brings it down to $0.20 per million input tokens and $1.20 per million output tokens. For context, that's a fifth of what Luna cost just yesterday.
The interesting part isn't just the number, it's how OpenAI got there. According to the company's own writeup, they used a third model, GPT-5.6 Sol, to optimize the infrastructure running everything else. Sol handled load balancing, sure, but it also went after the forward pass itself, the actual computation that turns input tokens into predictions. Idle GPU time from memory shuffling and clunky data layouts is apparently a bigger drag on cost than people assume, and Sol found ways to precompute, skip, or parallelize that wasted work.
More striking still: paired with Codex, Sol rewrote OpenAI's production kernels on its own. These are the low-level routines that execute the math underpinning every model call, written in Triton and Gluon, the GPU programming languages OpenAI maintains. Because GPT-5.6 was specifically trained to write and refine kernels in those languages, it could autonomously optimize the code that runs itself. That loop, model improving the infrastructure that serves models, shaved 20% off end-to-end serving costs.
The competitive fallout is immediate. Luna now undercuts Google's Gemini 3.1 Flash-Lite, which runs $0.25 input and $1.50 output, and it blows past Anthropic's Claude Haiku 4.5 at $1/$5, a model that was priced identically to Luna until today. Simon Willison, who runs a demo site on agent.datasette.io, said he'd already migrated it off Gemini Flash-Lite and onto Luna. When a hobbyist project switches providers within hours of a price cut, that's a decent signal the frontier of cheap-but-capable models just moved.
My take — AI-written commentary, not fact-checked reporting
I find it darkly funny that the pitch here is 'we used an AI to optimize the AI that serves the AI,' and it actually worked well enough to gut the competition's pricing overnight. This is the real AI arms race nobody talks about enough, not benchmark chasing but infrastructure cannibalization, and it's going to keep squeezing margins until running inference is basically free and the entire business model shifts to something else nobody's figured out yet.
Read more about this at: Simon Willison's Weblog
