What it is
GPT-6 Luna is OpenAI's most efficient current model, released on September 22, 2026 alongside GPT-6 Sol. It is aimed at focused, high-volume tasks. It reads text and images, writes text, and has a 1,050,000-token context, with input capped at 922,000 tokens and output at 128,000. Its knowledge cutoff is May 18, 2026.
The API price is $0.10 per million input tokens and $0.50 per million output tokens, with cached input at $0.01; Batch and Flex halve those rates. Unlike Sol and Astra, Luna can run with reasoning turned off (effort set to none), which keeps replies fast; the default is medium. Prompts over 272K input tokens are billed at twice the input rate and 1.5 times the output rate for the whole request.
It is available through the OpenAI API, in ChatGPT Work and Codex for paid users, and to Free and Go users through the ChatGPT desktop app.
Our review
Luna is the cheaper of the two budget models that matter, not the better one. OpenAI cut the price to $0.10/$0.50, half the input rate of GPT-5.6 Luna and about 58% less for output, and told VentureBeat the price is permanent. OpenAI says that at higher effort Luna matches GPT-5.6 Sol on factual accuracy at about a hundredth of the task cost, and reports 66.6% on DeepSWE 1.1.
Artificial Analysis puts numbers on the trade. At max effort Luna scores 38 on its Intelligence Index, 11th of 174 models, at $0.07 per task and 124 output tokens a second. Claude Haiku 5.5 scores 43 at $0.21 per task and 237 tokens a second, after generating 440M output tokens to Luna's 140M. Anthropic's own table shows a wider gap on agent work, 39.2% against 16.4% on Terminal-Bench 4.0 and 72.4% against 48.9% on OSWorld, though those are a rival's figures.
Its real advantage is range. Effort runs from none to max, so one model can serve instant classification and slower reasoning, and changing effort does not invalidate cached context. Long inputs stay cheap: a 200K-token contract costs about $0.02 to read on Luna, against about $0.10 on Haiku 5.5, whose rate jumps fivefold above 100K tokens.
The plumbing has edges. In Chat Completions, function calling only works with effort set to none; tools at any reasoning level need the Responses API, which stores responses for 30 days by default unless zero data retention is enabled. There is a single snapshot, gpt-6-luna, and no fine-tuning, so teams cannot pin a dated version or adapt the weights.
Where it earns its place
A marketplace trust-and-safety team labelling 20M listings a month
Classify every listing against a policy without a human first pass
Run Luna with effort none and structured outputs through the Batch API. At 800 input and 20 output tokens a listing, a month costs about $900 at Batch rates,, and batch jobs suit overnight queues.
A support platform sending 1M AI replies a month
Draft first replies to tickets at the lowest workable cost
A 1,500-token prompt and 300-token reply costs about $300 a month at list, or about $210 if a 1,000-token system prompt is cached. Reasoning tokens bill as output, so the default medium effort adds cost; test effort low or none first.
A legal-ops engineer summarising 200K-token contracts
Read whole agreements in one call without chunking
Luna bills a 200K-token prompt at the base $0.10 rate, about $0.02 of input per contract, because the long-context surcharge starts at 272K. The same input costs about $0.10 on Haiku 5.5, which reprices above 100K.
A European fintech whose data must stay in the EU
Use a cheap OpenAI model under EU data residency
EU data residency covers Luna on Standard, Fast, Flex and Batch processing, at a 10% regional surcharge. Zero data retention is available for Chat Completions and Responses, but not for Batch.
Before you commit
- Setup
- Call model gpt-6-luna on Chat Completions, Responses or Batch with an OpenAI API key. Set reasoning.effort explicitly, because the default is medium.
- Integrations
- Streaming, structured outputs, function calling, prompt caching and image input; the Responses API adds web search, file search, code interpreter, computer use, MCP and other hosted tools, each billed per call.
- Data and privacy
- OpenAI does not train on API data unless you opt in. Abuse-monitoring logs are kept up to 30 days, Responses API objects are stored for 30 days by default, and zero data retention is available for Chat Completions and Responses but not Batch. EU data residency is offered.
- Cost at scale
- 1M support replies of 1,500 input and 300 output tokens cost about $300 a month at list or $150 on Batch or Flex. Prompts over 272K tokens move the whole request to $0.20/$0.75, and Fast mode doubles the price.
- Limits
- 922K input and 128K output tokens. Default limits run from 5,000 requests and 2M tokens a minute on the Build tier to 30,000 requests and 180M tokens on Grow. At max effort Artificial Analysis measured a 90-second wait for the first token.
What's good
- $0.10/$0.50 per million tokens; $0.01 cached; Batch and Flex half price
- Reasoning can be turned off for low-latency calls
- $0.07 per Artificial Analysis task against $0.21 for Haiku 5.5
- Flat pricing up to 272K input tokens, where Haiku 5.5 jumps at 100K
What's not
- Scores 38 to Haiku 5.5's 43 on Artificial Analysis's index
- Function calling in Chat Completions only with effort set to none
- One unversioned snapshot and no fine-tuning
Best for
OpenAI-based apps doing classification, extraction or chat at very high volume.
Alternatives
GPT-6 Luna: common questions
- Is GPT-6 Luna worth it?
- Luna scores five points below Claude Haiku 5.5 on Artificial Analysis's index and costs a third as much per task, which is the whole decision in one line. We rate it 4.3 out of 5.
- How much does GPT-6 Luna cost?
- API: $0.10 per 1M input tokens, $0.50 per 1M output tokens; cached input $0.01; long prompts over 272K cost more.
- What are the downsides of GPT-6 Luna?
- Scores 38 to Haiku 5.5's 43 on Artificial Analysis's index; function calling in Chat Completions only with effort set to none; one unversioned snapshot and no fine-tuning.
- Who is GPT-6 Luna best for?
- OpenAI-based apps doing classification, extraction or chat at very high volume.
- How much does GPT-6 Luna cost for a team or in production?
- 1M support replies of 1,500 input and 300 output tokens cost about $300 a month at list or $150 on Batch or Flex. Prompts over 272K tokens move the whole request to $0.20/$0.75, and Fast mode doubles the price.
- Is GPT-6 Luna safe for company data?
- OpenAI does not train on API data unless you opt in. Abuse-monitoring logs are kept up to 30 days, Responses API objects are stored for 30 days by default, and zero data retention is available for Chat Completions and Responses but not Batch. EU data residency is offered.
- What are the best GPT-6 Luna alternatives?
- The closest alternatives we have reviewed are Claude Haiku 5.5, Gemini 3.5 Flash-Lite and GPT-6.1 Sol.
Answers checked Oct 22, 2026.
Make GPT-6 Luna? Show this rating on your site
