GPT 5.6 Luna

The GPT-5.6 model optimised for cost-sensitive, high-volume work, corresponding to the nano tier of earlier generations. OpenAI describes it as the fastest and most affordable model in the series, bringing strong capability at the lowest cost. It still reasons when asked, so quality degrades gracefully rather than falling off a cliff on harder inputs.

Luna is the cheapest and fastest tier in the family, and three weeks after launch OpenAI cut its price by 80%, to twenty cents per million input tokens and $1.20 per million output. The lab's framing for that move is worth reading twice: it says Luna delivers performance comparable to models that were frontier-class a year earlier at roughly six cents on the dollar per task and nearly nine times the speed, and that on its long-horizon professional evaluation Luna beats the strongest competing model at an estimated cost per task nearly 99% lower.

That is not a stripped-down completion engine. OpenAI reports Luna at 50.3% on Agents' Last Exam, 84.7% on Terminal-Bench 2.1, 62.7% on SWE-Bench Pro and 92.3% on GPQA Diamond, and says that on its coding-agent index Luna comes out ahead of the previous generation's flagship competitor. It reasons when asked, uses tools and completes multi-step workflows, which is what makes high-volume automation practical rather than merely cheap.

The workflow OpenAI itself sketches is the one to copy. Use the flagship to resolve uncertainty and settle the plan, then hand Luna the well-specified parts: implement the change, write and run the tests, evaluate the result. Deciding what to build is where frontier intelligence earns its rate; carrying out a decision that has already been made is where Luna's rate wins by two orders of magnitude.

The limits are specific and they are about capacity, not effort. On the lab's multi-round retrieval evaluation Luna scores 41.3% at both 256K to 512K and 512K to 1M tokens, against the flagship's 91.5% and 73.8% - so the million-token window is there, but do not rely on Luna to find a needle deep inside it. Abstract reasoning is the other cliff: 0.18% on the lab's newest novel-puzzle evaluation, against 7.78% for the flagship. Cybersecurity and self-improvement results fall off in the same way. Luna is for volume and for tasks whose shape you already know.