GPT Luna
by OpenAI · Model · Language & code models
promising
Cheap tier that cuts cost with similar quality, but makes more tool-call mistakes
GPT-6 Luna, cheapest tier at about 7 cents per task; free users get it in the desktop app
AccessAPIPrice$0.20 in / $1.20 out per M tokens
Why we say so
Datadog tested GPT-6 Luna in a live experiment: cost fell 44.3% against GPT-5.6 Luna with near-equal quality (3.86 vs 3.84 of 5), but it supplied invalid tool parameters more often, usually fixed on retry. Artificial Analysis found it more cost-efficient than Claude Haiku 5.5, while Anthropic claims Haiku wins on its own benchmarks. Microsoft's and Google's comparisons are the makers' own claims, while the Plotly bench shows Hy4 Flash beating Luna on data analytics.
Versions
| Version | Released | Rank |
|---|---|---|
| GPT-6 LunaLATEST promising Much cheaper than 5.6 with similar quality; tool calls need retries | Sep 27 | |
| GPT-5.6 Luna unproven Only appears as the baseline in Datadog's comparison |
Evidence
- Nov 17In Datadog's online experiment, GPT-6 Luna cut estimated cost 44.3% versus GPT-5.6 Luna with near-equal quality (3.86 vs 3.84 of 5)
- Nov 11Its decision model is priced at 10 cents per million input tokens with free output, and handles text and images
- Aug 28Commenters report Luna xhigh is very cheap and good; one finished a Quake-to-Raspberry-Pi port in OpenCode for about $0.40
- Aug 22GPT Luna (listed as GPT-5.6 Luna) scored 3,110 in 1.9 agent-days, roughly matching Sonnet 5 at 3,105
- Aug 17Reportedly cut by 80% on OpenRouter; usage jumped afterwards and it is competitive at the new price