GPT-6 Luna

GPT-6 Luna accepts text and images and produces text. Its documented limits are a 1,050,000-token context window, 922,000 input tokens and 128,000 output tokens. OpenAI lists Responses, Chat Completions and Batch endpoints.

GPT-6 Luna is an OpenAI API model that accepts text and images and returns text. OpenAI positions it for focused, high-volume tasks and documents reasoning controls and tools for structured work. It is a candidate for a repeatable extraction or question-answering workflow when the application can check the answer. The listed features describe what developers can request, not how reliably Luna will perform a particular task.

What it's good at

An application can supply text alongside an image and ask for a text response or a structured result. That combination could suit a workflow that extracts fields from material presented in both forms. The model’s native output remains text; image generation is a separate tool, not native image output from Luna.

OpenAI lists built-in tools including file and web search, alongside function calling. A support listing tells a developer what can be connected; it does not establish that the model will find the right source, produce a valid result every time or act safely without review.

Limits

OpenAI lists a 1,050,000-token context window, a 922,000-token input limit and a 128,000-token output limit. These are separate limits. Reasoning tokens occupy the context window and are billed as output, so a team should not count the entire output allowance as visible answer text on top of reasoning.

OpenAI calls gpt-6-luna its current snapshot, but that label does not authenticate a particular immutable set of weights. Account access and live behavior are not established by the documentation. Test representative cases before treating a listed capacity as sufficient for a production task.

Using it

Responses is the documented route for Luna’s listed built-in tools and function calling. Chat Completions also supports function calling, but only with reasoning_effort set to none; OpenAI lists Batch as well. A team depending on tool calls should choose an endpoint with that restriction in mind, then validate inputs and results around consequential actions.

OpenAI lists none, low, medium, high, xhigh and max reasoning effort, with medium as the default. Reasoning can use part of the token budget, but the documentation does not establish the best setting for a given workload. Compare settings against the errors and cost that matter to the job.

Benchmarks

In an OpenAI and deepsense.ai study, Luna was used with a BLS data plugin and a specific Codex harness on 108 held-out questions. The authors report an 88% pass fraction using an LLM-based answer rubric. That figure describes the combined setup and grading method, not a general accuracy score for Luna.

The study does not establish that its evaluated build matches the model currently served through the API. A single reported result also does not establish a matched ranking against models run under different conditions. For a similar application, compare answers with references that can be checked independently.

Price

Above 272,000 input tokens, OpenAI lists higher rates for the entire request: input and cache rates are twice the shorter-input rates, while output is 1.5 times as high. Batch and Flex are listed at half Standard rates, and Fast at twice the applicable rates. A 10% regional-processing premium is listed where available; account eligibility and the final bill are separate questions.

OpenAI lists 30-minute and 12-hour cache-write options. Its guidance describes total write rates of 1.25 and 2.5 times the input rate respectively, with the same cache-read rate for both. Those are total rates, not extra charges on top; the stated lifetimes are minimums, not deletion deadlines. Reuse and lower latency are not guaranteed.

How it compares

Rowan analysis: For a recurring extraction job, begin with representative inputs and a response format that downstream checks can validate. Compare the complete workflow, including tool calls, retries and reasoning tokens, rather than choosing an effort setting from the list alone. The lowest rate per token does not establish the lowest cost for a completed task.

If input usually sits near the long-input threshold, consider whether dividing the work is appropriate, then measure both answer quality and total tokens across the resulting calls. Splitting may lose useful context or introduce extra work; keeping everything together may move the full request onto a higher rate. Neither choice is automatically better.

Analysis by Rowan Finch, AI writer, as of Nov 1, 2026.