GPT-5.6 Luna Pro: A Request Setting Sold as a Model Name

If you see GPT-5.6 Luna Pro in a model list, the first thing to know is that there is no separate Luna Pro model. OpenRouter describes it as "the same underlying model as GPT-5.6 Luna, served with reasoning.mode set to pro". The name is a convenience for clients that pick behaviour by model ID. What you are actually choosing is an execution mode.

That matters because it changes how you should evaluate and pay for it. Older Pro models were priced as premium products. This one is priced exactly like Luna, and the extra cost arrives through token count instead of through the rate card.

What pro mode is

OpenAI introduced pro mode with the GPT-5.6 family on 9 July 2026, alongside persisted reasoning and the max effort level. Its reasoning guide says GPT-5.6 and GPT-6 models support two reasoning modes in the Responses API: standard, the default, and pro, "for difficult tasks that need more model work and can tolerate higher latency and token usage".

On OpenAI's API directly, you do not call a different model. You set a field:

{
  "model": "gpt-5.6-luna",
  "reasoning": { "mode": "pro", "effort": "medium" },
  "input": "..."
}

OpenAI documents the setting for the Responses API. OpenRouter's separate gpt-5.6-luna-pro ID exists so that a tool which can only change the model string can still reach it.

Same rate card, bigger token count

OpenRouter lists Luna Pro at $0.20 per million input tokens, $0.02 cached and $1.20 output at the time of writing — identical to plain Luna. That matches how OpenAI says pro mode bills: it "aggregates the model work performed to produce the final answer and bills those tokens at the selected model's standard" rates, and because it does more work, "token usage and cost" go up.

Compare that with the older Pro model IDs. GPT-5.5 Pro lists at $30 and $180 per million tokens, six times GPT-5.5. OpenAI says those IDs keep their behaviour and pricing. With pro mode the per-token price stays put and the whole premium is in how many tokens a request consumes, which you can see in the response's usage block. That is easier to budget, because you can measure it per request instead of trusting a multiplier.

The usual Luna rules still apply: prompts over 272K input tokens bill at twice the input rate and one and a half times the output rate for the full request. OpenRouter also lists a :batch variant at $0.10 and $0.60 — the same model and mode on a discounted asynchronous queue, which suits pro mode well, since pro requests are slow anyway.

Mode and effort are separate dials

This is the part people conflate. OpenAI says reasoning mode and reasoning effort are independent: mode selects the execution path, effort controls how much reasoning happens within it. Luna accepts six effort levels, from none to max, and defaults to medium in both modes.

That gives you at least three distinct ways to make Luna work harder, and they are not interchangeable:

• Standard mode, higher effort. More reasoning on the same path. high, xhigh and max sit here.
• Pro mode, same effort. A different execution path at the default level.
• Pro mode, higher effort. Both at once, and the most expensive combination.

OpenAI does not publish guidance on when pro beats a higher effort setting. Its general advice is to reserve xhigh for cases where your evaluations show a clear benefit, and to test whether max does better if you are already on xhigh. Treat pro mode the same way: an option to measure, not a default. Reasoning models explained covers why more thinking helps some tasks and is wasted on others.

Luna Pro, Luna at max, or a bigger model

The real decision is not Luna versus Luna Pro. It is which of these gets a hard task right most cheaply:

1. Luna at a higher effort in standard mode.
2. Luna in pro mode.
3. A larger model at standard settings — Terra at $2 and $12, or Sol at $4 and $20.

Luna is cheap enough that pro mode on it can still undercut a standard call to a bigger tier, but not automatically. Artificial Analysis found Luna at max effort far more verbose than Sol at the same setting, with a time to first token around two minutes. A small model doing a lot of work is still slow. If latency matters, a larger model at moderate effort may be the better buy even at a higher rate. If you want the same mode on the flagship, the GPT-5.6 Sol Pro guide covers it.

For most high-volume work, plain Luna is the right default. The GPT-5.6 Luna guide covers configuring it for cost. Pro mode is for the minority of requests where Luna's standard answer is wrong and you would rather spend more Luna tokens than switch models.

Measuring whether pro mode pays

Because the price lives in the token count, the test is straightforward:

1. Collect twenty tasks where plain Luna at medium gets the answer wrong.
2. Run each three ways: Luna at high, Luna Pro at medium, and Terra at medium.
3. Record correctness, total output tokens from the usage block, and wall-clock time.
4. Divide total cost by the number of correct answers for each configuration.

Pick whichever is cheapest per correct answer within your latency budget. How to benchmark LLMs on your own work covers building that task set so the result generalises beyond the twenty you picked.