At a glance
- OpenAI positions Luna for focused, high-volume tasks.
- Standard input and output rates are $0.10 and $0.50 per million tokens at verification.
- Measure cost per accepted result before replacing a production workflow.
Suppose ten thousand support tickets need an order number, an issue category and a human-follow-up flag. The output is small and the work repeats. The difficult cases are missing identifiers, several requests in one ticket and plausible guesses presented as facts.
Where Luna fits
OpenAI describes GPT-6 Luna as a model for focused, high-volume work. The identifier is gpt-6-luna; its page lists text and image input, text output and structured output support. Source 1 That makes extraction, classification and short summaries sensible trial tasks, rather than established production strengths.
The documented 1,050,000-token context and 128,000-token output ceiling describe capacity. Extracting three fields rarely requires sending an entire ticket archive.
Calculate rates, then measure acceptance
At verification, Standard API rates were $0.10 per million input tokens and $0.50 per million output tokens. Ten thousand requests using 1,000 uncached input tokens and 100 output tokens each would cost $1 for input and $0.50 for output: $1.50 before cache writes, tools, retries or other charges. These are vendor rates, not a Youjie quotation.
A project budget also needs repairs. If five hundred records require another call or manual correction, include that work. Compare total API expense and review time against the number of accepted records, rather than treating equal request counts as equal outcomes.
Include the awkward records
Build a labeled sample containing complete identifiers, missing identifiers and tickets with two separate requests. A refund plus an address change can reveal whether the model silently drops the second issue.
Specify null for missing values and an allowed list for categories. Count omissions, invented fields and invalid formats separately. Those categories suggest different fixes; a single accuracy score hides them. This article proposes an evaluation, not a completed Luna benchmark.
Check the tool interface
Luna’s documentation directs built-in tools and function calling to Responses. Chat Completions supports function calling only with reasoning_effort=none. Source 1 Replacing a model name alone may therefore break an existing tool workflow.
Inputs above 272,000 tokens also change the pricing tier. For bulk extraction, first constrain document length and missing-field behavior. Then test whether additional reasoning improves acceptance enough to justify its cost. Confirm the selected channel’s capabilities before making paid requests.
Sources & verification
Primary materials consulted for this article. Vendor measurements are not independent Youjie tests.
- GPT-6 Luna model documentation ↗OpenAI · Documentation · Checked Oct 19, 2026