GPT-6 Luna Review
Luna is worth testing for repeatable tasks whose outputs can be checked cheaply.
GPT-6 Luna launched on September 22, 2026 as OpenAI's efficient option for focused, high-volume work. At $0.10 input and $0.50 output per million tokens, its Standard token prices are one twentieth of Sol's. Our recommendation is to evaluate it for classification, extraction and short transformations, with a validator designed alongside the prompt.
Checked September 23, 2026. This launch review evaluates OpenAI's published specifications and reproducible cost calculations. We have not run a controlled benchmark of these models. Workload recommendations below are starting points for evaluation, not measured accuracy or speed results.
GPT-6 Luna review: who should use it?
Luna's strongest documented advantage is its price. It retains text/image input, reasoning controls and tool support rather than being a text-only completion endpoint. This makes it a practical candidate for the repeatable steps inside a larger application.
Use it where a failure can be detected and corrected: missing invoice fields, an invalid category or a broken JSON shape. Avoid deciding suitability from token price alone when a subtle mistake passes silently into a customer-facing answer or code change.
GPT-6 Luna specifications and context window
| Specification | GPT-6 Luna |
|---|---|
| API model ID | gpt-6-luna |
| Release date | 2026-09-22 |
| Context window (tokens) | 1,050,000 |
| Maximum input (tokens) | 922,000 |
| Maximum output (tokens) | 128,000 |
| Input / output | Text and images / text |
| Reasoning effort | none, low, medium, high, xhigh, max |
| Knowledge cutoff | 2026-05-18 |
Both models default to medium reasoning. A 1.05M context window is a capacity limit: the maximum input is 922K tokens, and the output allowance also needs room for reasoning. Matching context sizes do not establish matching recall or coding quality.
GPT-6 Luna API pricing
Direct OpenAI prices in USD per one million tokens, using Standard processing. These are not OmniaKey gateway quotes.
| Token category | Input up to 272K | Input above 272K |
|---|---|---|
| Uncached input | $0.10 | $0.20 |
| Cached input | $0.01 | $0.02 |
| Cache write | $0.125 | $0.25 |
| Output | $0.50 | $0.75 |
Above 272K input tokens, the entire request uses 2x input/cache rates and 1.5x output rates. The multiplier is not limited to the excess tokens. Exactly 272K remains in the lower bracket under the current wording.
OpenAI lists Batch and Flex at 50% of applicable Standard rates, Fast mode at 2x, and regional processing at a 10% premium where available. Processing options have their own eligibility and behavior; a price multiplier is not a latency guarantee.
How much do 1,000 Luna tasks cost?
| Illustrative workload | Token cost |
|---|---|
| 100K uncached input + 10K billed output | $0.015 |
| 300K uncached input + 10K billed output | $0.0675 |
| 1,000 jobs of 10K input + 1K billed output each | $1.50 |
A job with 10K input and 1K billed output costs 0.01 × $0.10 + 0.001 × $0.50 = $0.0015; 1,000 identical single-attempt jobs cost $1.50 in Standard token charges. At the documented 50% Batch rate the same token volume would be $0.75, before other charges and subject to Batch eligibility. These are specified workloads, not a promise that every document fits those counts.
These are calculations, not observed bills. Output includes billed reasoning tokens as well as visible text. The examples assume one attempt, no cache writes or hits, and exclude tool fees, storage, taxes and human review. Use actual usage fields when reconciling a run.
GPT-6 Luna performance and limitations
The official description emphasizes efficiency on focused tasks. It does not prove a particular accuracy level, tokens-per-second rate or equivalence to Sol. We have not run a first-hand Luna benchmark, so this assessment uses published functionality and cost math rather than a claimed performance score.
For extraction, measure missing fields, wrong values and invented values separately. A valid JSON document can still contain an incorrect invoice total. For classification, test ambiguous and out-of-distribution inputs as well as common labels. For code, start with a localized transformation and an existing test suite before handing Luna a repository-wide design decision.
Where Luna can fit in a production workflow
- Document processing: extract known fields and verify totals, required values and source references. Route uncertain cases to review.
- Classification and routing: restrict the label set, include an unknown path and measure false positives rather than accepting any syntactically valid answer.
- Bounded coding work: try small test additions or mechanical edits with executable acceptance criteria. Preserve a stronger-model fallback for unresolved cases.
A low-cost model can become expensive when it triggers repeated tool calls or human repair. Log the validation result for every attempt, cap retries and check a sample of accepted outputs for undetected errors. The Sol vs Luna comparison shows a hypothetical escalation budget.
API compatibility and tool calling
Both models support Responses and Chat Completions. For reasoning together with tools, use Responses API. Their model pages allow Chat Completions function calling only when reasoning_effort is none; endpoint support does not imply identical tool behavior.
On Responses, the documented reasoning.effort values are none, low, medium, high, xhigh, and max. Start with a fixed setting and raise it only when an acceptance test justifies the extra work. Function calling, structured outputs and image input are supported. Native output is text; using an image-generation tool is a separate operation.
curl --fail-with-body https://api.openai.com/v1/responses \
-H "Authorization: Bearer $OPENAI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "gpt-6-luna",
"reasoning": {
"effort": "medium"
},
"input": "Extract product names and quantities from this order. Return a JSON array: 3 notebooks and 2 pens."
}'
The request example and prices on this page refer to OpenAI directly. For OmniaKey, check the live model catalog for the exact route, supported features and current quote. OpenAI's release does not by itself verify a gateway route, a Codex picker or a ChatGPT plan entitlement.
Frequently asked questions
How much does GPT-6 Luna cost?
For Standard processing with up to 272K input tokens, the direct OpenAI rates per million tokens are $0.10 input, $0.01 cache read, $0.125 cache write and $0.50 output. Longer prompts use the higher rates shown above.
Can GPT-6 Luna write code?
It supports text generation, reasoning and tools, so it can be evaluated on coding tasks. Start with small, testable changes; its price and context window do not establish reliability on complex multi-file work.
Does Luna have the same context window as Sol?
Yes: both list 1,050,000 context tokens, 922,000 maximum input and 128,000 maximum output. These matching limits do not demonstrate matching long-context accuracy or speed.
Related reviews
- GPT-6 Sol review
- GPT-6 Sol vs Luna: which should you choose?
- GPT-6 Astra review
- View the model catalog