GPT 5.6 ಲೂನಾ ಪ್ರೊ

GPT-5.6 Luna running in pro reasoning mode -- the cheap tier given time to think. It is unusually good value on problems that need deliberation but not frontier knowledge, since you pay Luna rates for markedly better answers. Latency is correspondingly higher than plain Luna.

OpenAI ships one Luna and lets the caller decide how hard it thinks, with reasoning effort running from none through low, medium, high and xhigh to max. This entry is that model pinned to the deliberate end, which is an unusual thing to be able to buy: the price per token is the cheapest tier's price, and the only thing you spend more of is tokens and time.

The launch post says what the top of the dial does. Max gives the model more time than xhigh to reason, explore alternatives, run checks and revise its approach. That extra time is not decorative on a small model - in the preview post OpenAI reports Sol, Terra and Luna all showing strong improvements in capability as reasoning effort is raised, and the family's published numbers put Luna within a couple of points of the middle tier on several long-horizon professional tasks despite costing a tenth as much. On problems that need deliberation rather than recall, this is unusually good value.

It is also the setting where cost discipline matters most, because the multiplier lands on the token count rather than the rate. A deliberate answer from Luna can consume many times the tokens of a quick one, and it arrives correspondingly later. If a task has a known shape - a classification, a rewrite, a well-specified code change - plain Luna at a low effort is the right call and this is simply a slower way to pay for the same answer.

More thinking also cannot buy capacity the model does not have, and OpenAI's own tables say where that line is. Luna scores 41.3% on the lab's multi-round retrieval evaluation at every depth past 256K tokens, against 89.6% for the middle tier and 91.5% for the flagship, and 0.18% on its newest novel-puzzle evaluation. Those are limits of what the model can hold and what it has ever seen, not limits of how long it was allowed to think, so no effort setting will move them. Deep retrieval over a very long document and genuinely novel reasoning are escalation cases, not effort cases.