What is GPT-6 Luna?
GPT-6 Luna is OpenAI’s low-cost GPT-6 model for focused, high-volume tasks. Released September 22, 2026 alongside GPT-6 Sol, it supports the same 1.05-million-token context window, image input, reasoning controls and broad Responses API tool set at much lower token prices.
The model costs $0.10 per million input tokens and $0.50 per million output tokens at Standard rates. That makes Luna attractive for routing, extraction, classification, support, repetitive coding work and agent subtasks where cost and throughput matter more than maximum reasoning depth.
QUICK VERDICT: GPT-6 Luna is OpenAI’s most efficient current model for focused, high-volume work. Its $0.10 input and $0.50 output rates are unusually low for a reasoning-capable multimodal model, but complex or ambiguous work should be escalated to GPT-6.1 Sol or Astra when Luna does not reliably clear the acceptance threshold.
Field | Verified value |
|---|---|
Provider | OpenAI |
Release date | September 22, 2026 |
License | Proprietary |
API model ID | gpt-6-luna |
Context window | 1,050,000 tokens |
Maximum input | 922,000 tokens |
Maximum output | 128,000 tokens |
Modalities | Text and image input; text output |
Knowledge cutoff | May 18, 2026 |
Reasoning effort | None, low, medium, high, xhigh and max; medium is default |
Base API pricing | $0.10 input, $0.01 cached input and $0.50 output per million tokens |
Best fit | Focused high-volume workloads and inexpensive agent subtasks |
Last verified | October 2, 2026 |
What changed from GPT-5.6 Luna
GPT-6 Luna cuts token prices in half relative to GPT-5.6 Luna’s promotional rates while improving professional work, coding, computer use, factuality and communication style. It also gains GPT-6 prompt-caching improvements and the same broad tool surface as the larger family models.
Area | GPT-6 Luna change |
|---|---|
Input price | Reduced from $0.20 to $0.10 per million tokens |
Output price | Reduced from $1.20 to $0.50 per million tokens |
Professional work | Higher reported AutomationBench results at lower task cost |
Coding | Provider-reported gains on DeepSWE and related evaluations |
Computer use | More efficient OSWorld performance than the previous Luna tier |
Communication | Clearer, shorter and less jargon-heavy style derived from GPT-6 Astra work |
The practical question is not whether GPT-6 Luna is newer. It is whether the change improves accepted-task quality, speed and cost in the exact workflow. Keep prompts, tools, source material and scoring criteria fixed when comparing it with GPT-5.6 Luna.
Capabilities and best use cases
Luna supports text and image inputs plus web search, file search, code execution, hosted shell, computer use, MCP and other Responses API tools. Its advantage is not a unique tool set. It is the ability to use those tools at a much lower token rate when the workload is bounded and well specified.
Capability | Practical use |
|---|---|
Classification | High-volume routing, labeling and structured decisions |
Extraction | Fields, entities and summaries from text, documents and images |
Customer support | Drafting and tool-assisted resolution for routine cases |
Coding subtasks | Focused fixes, tests, transformations and repository assistance |
Agent routing | Cheap first pass before escalating difficult cases to Sol or Astra |
Long context | Large inputs when retrieval or selective processing is still carefully evaluated |
The strongest documented fit is cost-sensitive, high-volume classification, extraction, support and agent subtasks. A capable base model can still fail when retrieval is weak, tools are over-permissioned, instructions conflict or the system lacks verification. Evaluate the full application rather than the model in isolation.
Benchmark evidence
OpenAI reports 66.6% on DeepSWE v1.1 at maximum effort, along with gains over GPT-5.6 Luna on AutomationBench and OSWorld. The company notes that Luna can approach more expensive older models on selected tasks. These are provider comparisons and should be validated against the exact quality threshold of a high-volume workflow.
Evaluation | Reported result | Evidence boundary |
|---|---|---|
DeepSWE v1.1 | 66.6% at max effort | OpenAI-reported coding-agent result |
AutomationBench | 5.4 points above GPT-5.6 Luna at high effort | Provider-reported business-workflow gain |
OSWorld 2.0 offline | Exceeds GPT-5.6 Sol medium at max effort | Provider comparison at about one tenth of the cited cost |
Internal factuality | Substantial improvement over GPT-5.6 Luna | Provider evaluation on difficult flagged conversations |
Collaboration style | Shorter and clearer responses | Qualitative provider assessment rather than a benchmark score |
Benchmark scores depend on model version, reasoning effort, sampling, tools, scaffolding and the exact dataset release. Do not combine scores from different harnesses into a synthetic ranking. Use public results to choose candidates, then test those candidates on frozen tasks from the intended workflow.
Artificial Analysis Intelligence Index
Artificial Analysis independently measures model intelligence, speed and cost. GPT-6 Luna scores 38 on the Artificial Analysis Intelligence Index at max effort. Lowest cost per Intelligence Index task of any tracked model at launch.
Metric | Value |
|---|---|
Intelligence Index score | 38 |
Effort setting | max |
Source | Artificial Analysis (independent) |
Last verified | October 2, 2026 |
The Artificial Analysis Intelligence Index combines scores across mathematics, reasoning, coding, instruction following and language tasks. The scale is not a percentage: the index reflects relative position across the models they track, not a share of correct answers. Compare scores only within the same effort setting and the same index version.
Pricing and access
GPT-6 Luna is available through the OpenAI API, ChatGPT Work and Codex. Free and Go users received desktop-app access at launch. The model supports the Responses API and Chat Completions, with function calling in Chat Completions restricted to reasoning effort set to none.
Luna is most valuable as part of a routing system. Send routine cases to Luna, define observable escalation criteria and route uncertain or consequential cases to GPT-6.1 Sol, GPT-6 Astra or a qualified human reviewer.
Usage or access item | Current value |
|---|---|
Standard input | $0.10 per million tokens |
Cached input | $0.01 per million tokens |
Cache writes | $0.125 per million tokens |
Standard output | $0.50 per million tokens |
Batch and Flex | 50% below Standard rates |
Fast mode | 2 times Standard rates |
Prompts above 272K | 2 times input and cache; 1.5 times output for the full request |
Token price is only one part of total cost. Include cache behavior, long-context multipliers, tool calls, retries, wall time, failed-task recovery and human review. The useful comparison is cost per accepted result, not price per million tokens in isolation.
Limitations and deployment risks
- Luna is not the best choice for the hardest reasoning, coding or judgment-heavy work.
- Low token prices can hide high retry, tool and review costs if task quality is below the acceptance threshold.
- Prompts above 272,000 tokens trigger long-context multipliers across the whole request.
- Maximum-effort benchmark results may not reflect the latency or cost of a high-volume deployment.
- Long context should not replace retrieval, document selection and explicit evidence requirements.
- Tool-using agents need escalation rules, permission boundaries and failure monitoring.
- A routing system must measure false escalations and missed difficult cases, not only average cost.
High-stakes legal, medical, financial, scientific and security work needs qualified review. Log the exact model version and configuration, separate untrusted content from system instructions, restrict tools to the minimum required scope and require approval before irreversible or externally visible actions.
How to evaluate GPT-6 Luna
Build a frozen evaluation set of 20 to 50 real tasks. Include routine work, difficult edge cases, adversarial inputs, long-context examples and cases where the correct behavior is to stop or escalate. Compare GPT-6 Luna with GPT-5.6 Luna and at least one neighboring model using equivalent tools and source material.
Test area | What to record |
|---|---|
Task completion | Pass or fail against a written acceptance rubric |
Reliability | Repeated-run success rate, variance and silent failures |
Factuality | Unsupported claims, source use, quotations and citation accuracy |
Tool use | Wrong calls, retries, recovery and permission-boundary failures |
Long context | Retrieval accuracy, instruction retention and cost at realistic lengths |
Efficiency | Wall time, input, output, cached tokens, tool charges and review time |
Safety | Prompt injection, sensitive-data handling and irreversible-action controls |
Choose the least expensive configuration that clears the acceptance threshold with a safety margin. Re-run the suite when the model, system prompt, effort level, retrieval layer, tool definitions or approval policy changes.
Who should use it?
Situation | Recommendation |
|---|---|
Strong fit | cost-sensitive, high-volume classification, extraction, support and agent subtasks |
Pilot first | Long-running agents, large contexts, computer use and workflows with several tools |
Escalate | Ambiguous or consequential work that does not reliably clear the evaluation threshold |
Avoid unsupervised use | Irreversible actions, sensitive data or high-stakes decisions without monitoring and approval |
A migration should be driven by measured outcomes. Keep GPT-5.6 Luna available during the pilot, record where each model succeeds or fails, and use routing when different task classes have different quality and cost requirements.
Frequently asked questions
Is GPT-6 Luna open source?
No. GPT-6 Luna is proprietary. Access, serving behavior and lifecycle decisions are controlled by OpenAI and supported distribution partners.
How much does GPT-6 Luna cost?
$0.10 per million tokens. Review the full pricing table above because cached input, long context, processing mode and platform can materially change the total.
What is GPT-6 Luna best used for?
Its strongest documented fit is cost-sensitive, high-volume classification, extraction, support and agent subtasks. Start with a supervised pilot and retain human sign-off for consequential work.
Should I migrate from GPT-5.6 Luna?
Only after a side-by-side evaluation. Measure accepted-task quality, total cost, latency, output style, tool reliability and migration engineering. A newer model is not automatically the better operational choice.
How should benchmark evidence for GPT-6 Luna be interpreted?
Use provider results to identify promising workloads, then reproduce the comparison with the exact model configuration, tools and acceptance criteria that matter to the deployment. Do not infer a site ranking from benchmark claims gathered under a different harness.
Related model guides
Browse the Best AI Models directory.
Compare with the GPT-6 Astra guide.
Compare with the GPT 5.6 Sol guide.
Compare with the GPT 5.5 guide.
Official sources and update policy
OpenAI’s GPT-6 Sol and Luna announcement.
GPT-6 Luna API model documentation.
Checked October 2, 2026. We update this guide when OpenAI changes the specification, pricing, access, safety documentation or model lifecycle. Vendor benchmarks are attributed and are not presented as independent testing. No vendor payment or affiliate relationship determined inclusion.