ModelsGPT-6 Luna Pro: Buying Depth Without Moving Up a Tier
GPT-6 Luna Pro runs Luna in pro reasoning mode at Luna's own rates. How it differs from raising effort, when it beats stepping up to Sol, and how to measure it.
When a cheap model gets a task slightly wrong, the usual fix is to send that task to a more expensive model. GPT-6 Luna Pro is a different option: keep the cheap model, and make it work harder on that request. Whether that is a better trade than escalating depends on the task, and it is cheap enough to find out.
Luna Pro is not a separate model. It is GPT-6 Luna, released by OpenAI on 22 September 2026, served with reasoning mode set to pro. OpenRouter lists it under its own id, openai/gpt-6-luna-pro, at the same $0.10 input and $0.50 output per million tokens as standard Luna at the time of writing.
What the pro suffix changes
OpenAI's reasoning guide describes two execution modes for GPT-5.6 and GPT-6 models in the Responses API. standard is the default. pro is for difficult tasks that need more model work and can tolerate higher latency and token usage. Pro mode bills at the model's normal rates, so the extra cost comes entirely from the extra tokens it generates.
The guide is also explicit that mode and reasoning effort are independent. Effort, from none to max on Luna, controls how much reasoning the model applies. Mode decides whether it runs the standard or pro execution path. So Luna Pro is not the same as Luna at max effort, and it is worth testing both.
Everything else is inherited from Luna: a 1,050,000-token window with up to 922,000 tokens of input, 128,000 tokens of output, text and image input, and the same long-context pricing step for prompts above 272K tokens. The GPT-6 Luna guide covers those basics.
Why the per-token price is misleading here
On a pricing page, Luna and Luna Pro look identical. On an invoice they will not be. If pro mode doubles or triples the output tokens for a request, the request costs two or three times as much. OpenAI does not publish a typical multiplier, and it varies with the task, so the only reliable number is the one you measure.
What makes Luna Pro interesting is the base it starts from. Luna's output rate is $0.50 per million tokens. GPT-6 Sol's is $10, twenty times higher. Even a generous pro-mode multiplier on Luna leaves a lot of room before it reaches Sol's cost for the same task. That is the whole case for this variant: depth bought at the bottom of the price list.
The token cost of reasoning models explains why hidden reasoning output is the part of the bill people underestimate.
Luna Pro versus simply using Sol
The honest comparison is not with standard Luna but with GPT-6 Sol, because that is what you would otherwise escalate to.
- Luna Pro tends to suit tasks that are bounded but fiddly: a tricky extraction schema, a classification with many overlapping labels, a short piece of logic that needs checking. More deliberation on a small model can close the gap.
- Sol tends to suit tasks that need knowledge or capability the small model does not have: multi-file code changes, unfamiliar libraries, long agent loops. Thinking longer does not supply missing ability.
- Latency cuts against Luna Pro. Pro mode takes longer by design. If the request is user-facing, standard Sol may return a correct answer sooner than Luna Pro does.
A useful way to tell the two cases apart: look at how Luna fails. If its wrong answers are careless, such as a missed field or an inconsistent label, pro mode is a good bet. If they are confidently wrong about how something works, move up a tier.
Wiring it into an escalation ladder
Luna Pro fits naturally as the middle rung of a three-step ladder:
- Send the request to standard Luna at low or medium effort.
- If a validator rejects the result (schema check, test failure, confidence threshold), retry on Luna Pro.
- If that also fails, escalate to Sol.
Because most requests never leave step one, the average cost stays close to Luna's, and the share that reaches Sol shrinks. Track how many requests stop at each rung; escalation rate as a cost metric explains why that number decides your real cost better than any list price.
For offline work, the catalogue also lists openai/gpt-6-luna-pro:batch: the same model and mode on an asynchronous endpoint at half price, $0.05 input and $0.25 output per million tokens at the time of writing. Pro mode's latency stops mattering when nobody is waiting, so batch and pro pair well.
Measuring whether it earns its place
Take fifty requests where standard Luna got the answer wrong. Run each through Luna at max effort, Luna Pro at medium, and Sol at medium. For each, record whether the answer is correct, the output tokens used, and the time to respond.
Then compute cost per corrected answer for each option. If Luna Pro fixes most of the failures at a fraction of Sol's cost, it belongs in your routing. If it fixes few of them, the failures were about capability rather than effort, and you can drop the middle rung and escalate straight to Sol.
Common questions
Is GPT-6 Luna Pro more expensive than GPT-6 Luna?
Not per token: both list at $0.10 input and $0.50 output per million tokens. Each request usually costs more because pro mode generates more reasoning tokens, so measure the real cost on your own tasks.
Should I use GPT-6 Luna Pro or GPT-6 Sol?
Use Luna Pro when Luna's errors look like carelessness on bounded tasks, and Sol when the task needs more capability, such as multi-file coding or long agent loops. Latency-sensitive requests usually favour standard Sol.
What is the difference between pro mode and max reasoning effort?
They are separate settings. Effort controls how much reasoning the model applies, while mode selects standard or pro execution. OpenAI documents them as independent, so test both rather than assuming one includes the other.
