Glonce

OpenAI cuts GPT-5.6 Luna price 80%, Terra by 20%

modelsThe Glonce editorsSimon Willison🔥76Glonce rating76Impact72Novelty52Relevance100Surprise72Players90Credibility68

OpenAI cut prices across its GPT-5.6 model family on July 30, 2026. GPT-5.6 Terra got a 20% reduction, while GPT-5.6 Luna, the smallest and cheapest model in the line, dropped by 80%. Luna now costs $0.20 per million input tokens and $1.20 per million output tokens.

OpenAI attributes the savings to work done by another model in the family, GPT-5.6 Sol. In a company blog post titled "How GPT-5.6 fuses frontier intelligence with frontier efficiency," OpenAI describes using Sol to optimize load balancing and, more strikingly, to optimize inference itself. Sol was set loose on the model's forward pass, the computation that turns inputs into next-token predictions, and found work that could be precomputed, skipped, or run in parallel instead of sequentially. Using Codex, Sol autonomously rewrote and optimized OpenAI's production kernels, the low-level code that carries out the model's core mathematical operations. OpenAI says this was possible in part because GPT-5.6 itself had been trained to write and improve kernels in Triton and Gluon, two open-source GPU programming languages that OpenAI maintains. Combined with other kernel-level advances credited to Sol, the work cut end-to-end serving costs by 20%.

The Luna price cut resets the field for cheap models. Commentator Simon Willison, writing on his link blog, notes that Luna is now cheaper on input than Google's Gemini 3.1 Flash-Lite, priced around $0.025 per million input tokens and $1.50 per million output tokens. Willison also compares Luna to Anthropic's cheapest current model, Claude Haiku 4.5, priced at $1 per million input tokens and $5 per million output tokens: Luna's new input price is one-fifth of Haiku 4.5's, where the two used to be priced the same. Willison says he switched his own demo site, agent.datasette.io, from running on Gemini 3.1 Flash-Lite over to Luna as a direct result of the price change.

Key facts

  • OpenAI cut GPT-5.6 Terra's price by 20% and GPT-5.6 Luna's price by 80% on July 30, 2026.
  • Luna now costs $0.20 per million input tokens and $1.20 per million output tokens.
  • OpenAI says GPT-5.6 Sol optimized the forward pass and, using Codex, autonomously rewrote production kernels written in Triton and Gluon, cutting end-to-end serving costs by 20%.
  • Luna's new input price undercuts Google's Gemini 3.1 Flash-Lite (about $0.025/$1.50 per million tokens) and is now one-fifth of Anthropic's Claude Haiku 4.5 input price ($1), where the two previously matched.
  • Simon Willison switched his agent.datasette.io demo site from Gemini 3.1 Flash-Lite to Luna after the cut.

Why it matters

The size of the Luna cut, 80% in one move, is unusual even by the standard of a market where model prices fall constantly. What makes it notable is the stated mechanism: OpenAI says the saving came from a smaller model, GPT-5.6 Sol, rewriting the production kernels that run inference, rather than from a simple margin cut. If that holds up, it points to AI companies using their own models to compress the cost of serving other models, which would keep pushing the price of cheap-tier inference down independent of hardware gains.

Who it affects

Anyone building on the cheapest tier of frontier-adjacent models: developers of high-volume, cost-sensitive applications such as demos, agents, and background automation who previously picked between Gemini 3.1 Flash-Lite and Claude Haiku 4.5 on price. Simon Willison is a concrete example: he moved his own agent.datasette.io demo from Gemini's cheap tier to Luna specifically because of this price change.

How to use it

Luna is priced at $0.20 per million input tokens and $1.20 per million output tokens, available through OpenAI's standard API. Terra's exact new price is not given in the source, only the 20% reduction. No new sign-up or plan change is described; this is a standard API price update, not a new tier or product.

How solid is it

The account comes from Simon Willison's link blog, which relays and comments on OpenAI's own announcement blog post; the underlying efficiency claims (kernel rewrites via Sol and Codex, the 20% serving-cost reduction) are OpenAI's characterization of its own work and are quoted directly from that post rather than independently verified. The pricing figures for Luna and the comparison to Anthropic's Haiku 4.5 are stated plainly; the Gemini 3.1 Flash-Lite figure is given with less precision in the source.

Risks and caveats

This is a secondary account of a primary OpenAI announcement, not an independent benchmark of the cost or quality tradeoffs at Luna's new price point. The claimed engineering process, a model autonomously rewriting its own production kernels, is described by OpenAI itself and has not been externally audited. Prices in this segment move quickly and may already have shifted by the time this is read.

“With Codex, GPT-5.6 Sol autonomously rewrote and optimized our production kernels, the core code that executes the mathematical operations that make up the model.”

— OpenAI, "How GPT-5.6 fuses frontier intelligence with frontier efficiency"