model-release
GPT-6 Luna Release Date: Out Sept 22, 2026 — OpenAI's Most Efficient Model, Rollout by Channel
Jessie
COO
September 20, 2026
Updated on September 23, 2026
7 min read
GPT-6 Luna was released on September 22, 2026. OpenAI calls it "our most efficient model for focused, high-volume tasks" and prices it at half the GPT-5.6 Luna list rate. OpenAI's API changelog lists GPT-6 Sol and GPT-6 Luna as released that day through the Responses and Chat Completions APIs. The official model reference documents the model ID
gpt-6-luna, text and image input, text output, a 1,050,000-token context window (922,000 maximum input), 128,000 maximum output tokens and a knowledge cutoff of May 18, 2026. OpenAI positions it for well-defined, high-frequency tasks.If you process a queue of documents or classifications, the date answers one question and opens three more: which channel serves the model, what it costs on your route, and whether it meets your accuracy and throughput targets. This article tracks the rollout. GPT-6 Luna is live on EvoLink since September 23, 2026 as
gpt-6-luna; the GPT-6 Luna API page owns the current rates, request examples and supported features; GPT-6 Luna vs GPT-5.6 Luna owns the migration evaluation.What is confirmed about the Luna rollout?
| Milestone | Status checked September 23, 2026 | Evidence |
|---|---|---|
| Official release date | September 22, 2026 | OpenAI API changelog entry naming GPT-6 Sol and GPT-6 Luna |
| Official API reference | Live: model ID gpt-6-luna, Responses and Chat Completions | Official model reference |
| Official rate | Published: Standard, long-context (above 272K input tokens), Batch, Flex and Fast rows | Official pricing page |
| Codex | Rolling out from September 22 for Plus, Pro, Business, Enterprise and Edu users; eligibility is per plan and account | OpenAI announcement; staged per account |
| ChatGPT | Rolling out from September 22 in ChatGPT Work for Plus, Pro, Business, Enterprise and Edu; Free and Go users can try GPT-6 Luna in the desktop app | OpenAI announcement |
| EvoLink availability | Callable since September 23, 2026 as gpt-6-luna; current rates on the API page | EvoLink route activation |
Each row moves on its own evidence; one channel's status says nothing about another's. The EvoLink rate and request examples live on the product page, where they are maintained with the route.
What GPT-6 Luna is built for, in OpenAI's words
OpenAI's model page positions GPT-6 Luna as "our most efficient model for focused, high-volume tasks": summarization, extraction, classification and straightforward question answering, while keeping the 1.05M-token context, 128K output, selectable reasoning effort and the Responses API tool set. The launch announcement adds the claims below. They are vendor-reported; none has been reproduced on EvoLink.
| OpenAI's claim | Reported figure | What it does not prove |
|---|---|---|
| Price | Half the GPT-5.6 Luna list rate; the cheapest GPT-6 tier | Your EvoLink rate, or your cost per accepted record |
| Factuality | At higher reasoning effort, matches GPT-5.6 Sol's factuality at about one-hundredth of its task cost | Field accuracy on your documents and labels |
| Agent honesty | 2.8% deception rate in OpenAI's internal test, down from 9.5% for GPT-5.6 Luna | Behaviour under your schema and tool constraints |
| Caching | Higher default prompt-cache hit rates and a 90% discount on cached input reads | Cache hit rates on your own prompts and route |
Sources: OpenAI announcement and model page, read September 23, 2026; secondary coverage by TechCrunch and VentureBeat. The upgrade guide turns these claims into a record-level acceptance set.
Distinguish GPT-6 Luna from the existing GPT-5.6 Luna
OpenAI now documents two Luna models. GPT-5.6 Luna is
gpt-5.6-luna, positioned for cost-sensitive, high-volume work. GPT-6 Luna is gpt-6-luna, with a newer knowledge cutoff (May 18, 2026) and the same 1,050,000-token context and 128,000 maximum output. The tier name carries over; the identity does not.Check the full identifier before interpreting a dashboard label, search result or release summary. An article that compares GPT-6 Astra with GPT-5.6 Luna contains "GPT-6" and "Luna" without describing GPT-6 Luna. Do not replace an existing queue's model field with a slug from a web page; keep the current route explicit and add the new one as a separate configuration entry after the provider you use documents it.
Which release signals matter to a batch-processing team?
This is an illustration of the verification process, not a completed GPT-6 Luna benchmark.
For repeated workloads, access is only the start of the operational check. What each signal establishes:
- The changelog entry establishes the name and date. It says nothing about your account, plan or gateway.
- The model reference establishes upstream capabilities: reasoning effort from
nonetomax, streaming, structured outputs, and function calling on the Responses API at any effort but on Chat Completions only withreasoning_effortset tonone. A gateway still needs its own implementation check. - A successful request establishes access for one route and account at that moment. It does not prove sustained throughput at your concurrency.
- A priced, observed trial establishes whether your queue meets accuracy, latency and budget targets. A lower list price per token does not answer whether retries consume the saving.
Prepare the queue before planning the migration date
Start with a small representative sample from work you already understand: ordinary records, missing information, long inputs, ambiguous labels and records your current workflow escalates. Save the expected fields or decisions and the rules for allowing an empty value or an explicit refusal.
Then record three views of the GPT-5.6 Luna baseline:
- Correctness: accepted records divided by all attempted records, with schema errors separated from wrong values.
- Completion: accepted records per minute and end-to-end P95 time, including queue wait and retry delay.
- Spend: the full pipeline bill per accepted record, including failed attempts and any escalation to another model.
If your pipeline calls tools with reasoning enabled on Chat Completions, decide now whether the GPT-6 Luna trial runs on the Responses API or with
reasoning_effort: none. The Luna upgrade guide adds a field-level test matrix and rollback criteria. For present-day choices, see the GPT-5.6 product page and GPT collection.What will be updated next
EvoLink activated GPT-6 Luna on September 23, 2026, one day after the OpenAI release; the two dates are recorded separately. Prices and request details stay on the product page. Evaluation results go into the comparison guide only after a documented trial with its sample, settings and limitations. Keep your existing queue on its measured route until your own sample clears the switch, and check current GPT-6 Luna rates before sizing the trial.
FAQ
When was GPT-6 Luna released?
September 22, 2026. OpenAI's API changelog lists GPT-6 Sol and GPT-6 Luna as released that day.
What is the GPT-6 Luna model ID?
gpt-6-luna, also the snapshot ID in the official model reference. EvoLink uses the same ID.Is the documented Luna model already GPT-6?
There are now two:
gpt-5.6-luna and gpt-6-luna. A reference to "Luna" without a generation number can mean either; check the identifier and the knowledge cutoff.Did GPT-6 Luna launch with another GPT-6 model?
Yes. GPT-6 Sol was released the same day. Each has its own model reference and pricing rows; both are callable on EvoLink.
Is GPT-6 Luna available on EvoLink?
Yes. Since September 23, 2026 you can call
gpt-6-luna on EvoLink with your existing API key, on Chat Completions and Responses. Current rates are on the GPT-6 Luna API page.Can I set a queue migration date now?
Yes for the evaluation window, since the model is callable on EvoLink today; keep the production cut-over conditional on measured accuracy, throughput and cost per accepted record on your own sample.
