OpenAI has cut GPT-5.6 Luna’s price by 80%. The model now costs $0.20 per million input tokens and $1.20 for output (source: OpenAI, July 30, 2026). The largest model in the family, Terra, took a 20% cut.

For those building agents, the math of model selection changes. Until yesterday the trade-off was clear: frontier model for hard tasks, compact model for the rest, because the frontier was too expensive to use everywhere. Now Luna costs less than Gemini 3.1 Flash-Lite and a fifth of Claude Haiku 4.5 on input. The boundary between capable and economical model is blurring.

The cut is made possible by GPT-5.6 Sol, which OpenAI used to rewrite inference kernels (the code that runs model computations on GPU) in Triton and Gluon, two open-source GPU languages from the lab. End-to-end serving costs 20% less, and OpenAI passed the savings on to pricing.

As we reported on July 25, model selection is shifting from benchmarks to cost per verified output. With Luna at this price, many tasks that ended up on a compact model to save money can move back to a frontier without changing the bill.

In detail

Luna’s price before the cut matched Claude Haiku 4.5 on input: $1 per million tokens. With the new pricing at $0.20, input costs drop five-fold. On output the gap is less extreme ($1.20 vs Haiku 4.5’s $5), but still wide. The comparison with Gemini 3.1 Flash-Lite is stark: Luna costs less, despite being a model from OpenAI’s frontier family.

What makes the cut possible. OpenAI attributes the savings to GPT-5.6 Sol, the model in the family optimized for deep reasoning. Sol worked on two fronts: load balancing and, more importantly, rewriting inference kernels. Kernels are the code that executes the model’s mathematical operations on GPU. When individual operations are already fast, what slows things down is data movement in memory, synchronization between operations, and inefficient data layouts. Sol found computations that could be precomputed, avoided, or parallelized, and rewrote production kernels in Triton and Gluon, two open-source GPU languages maintained by OpenAI. The stated result is a 20% reduction in end-to-end serving costs.

What changes for builders. If you have an agent using a compact model to keep costs down, do the math again. A research agent processing hundreds of documents, email triage, an assistant answering FAQs: tasks where the compact’s quality was “good enough” but the frontier was too expensive. With Luna at $0.20 per million input tokens, reading a long document (say 10,000 tokens) costs a fraction of a cent. The quality difference between a frontier and a compact, on tasks requiring understanding and judgment, might be worth that spend.

A concrete signal: a demo agent on Datasette that ran on Gemini 3.1 Flash-Lite was migrated to Luna the day of the announcement.

Where to be careful. Published prices are official OpenAI rates. On third-party providers (Bedrock, Azure) pricing can differ, and prompt caching further changes the math. The 20% cut in serving costs is an OpenAI claim, not a number verified by third parties. And price per token isn’t the only cost of a production agent: latency, call count, retries, and orchestration overhead matter as much as the list price.

To decide if Luna makes sense for your case, the method stays the same: same task, same test cases, results table with cost and quality. The playbook to do it in a quarter hour is here.