GPT-6 Luna

What is disclosed

GPT-6 Luna came out on 22 September 2026, on the same day as GPT-6 Sol.¹¹1OpenAI, API changelog, entries of 22 and 25 September 2026. https://developers.openai.com/api/docs/changelog OpenAI's GPT-6 guide lists it as the fastest and most cost-effective model of the family.²²2OpenAI, Using GPT-6 guide, read 29 September 2026. https://developers.openai.com/api/docs/guides/latest-model

Luna keeps the full 1,050,000-token window of GPT-6 Astra and its 128,000-token output limit.³³3OpenAI, GPT-6 Luna model page, read 29 September 2026. https://developers.openai.com/api/docs/models/gpt-6-luna Like the other two, it works on text and images and answers only in text. Its knowledge runs later than Astra's, to 18 May 2026 against 30 April.³³⁴⁴4OpenAI, GPT-6 Astra model page, read 29 September 2026. https://developers.openai.com/api/docs/models/gpt-6-astra

The system card says Luna uses the same types of data and training as Astra.⁵⁵5OpenAI, GPT-6 Astra System Card, HealthBench section and appendix on GPT-6 Sol and GPT-6 Luna, added 22 September 2026. https://deploymentsafety.openai.com/gpt-6-astra How large it is and how it is built are not published.

What is new

Its predecessor, GPT-5.6 Luna, came out on 9 July 2026. OpenAI placed that model roughly at the nano tier of earlier GPT-5 families.⁶⁶6OpenAI, GPT-5.6 Luna model page, read 29 September 2026, with the release date from the API changelog entry of 9 July 2026. https://developers.openai.com/api/docs/models/gpt-5.6-luna and https://developers.openai.com/api/docs/changelog The knowledge cutoff moves from 16 February to 18 May 2026.⁶⁶³³

Standard input costs $0.10 per million tokens, half the $0.20 of GPT-5.6 Luna.⁷⁷7OpenAI, API pricing, Standard table, read 29 September 2026. https://developers.openai.com/api/docs/pricing. Arithmetic: 0.10 / 0.20 = 0.5; 0.50 / 1.20 = 0.42; 0.01 / 0.10 = 0.1. Output costs $0.50 against $1.20, about 42% of the older model's rate.⁷⁷

On HealthBench Professional, length-adjusted, Luna scores 60.8, or 5.1 points above GPT-5.6 Luna.⁵⁵ On the plain HealthBench and on HealthBench Hard it scores slightly lower than its predecessor. OpenAI ties the drop to answers about 35% shorter.⁵⁵

OpenAI treats Luna as High capability in cybersecurity and in biology and chemistry.⁵⁵ It needed no separate Critical testing, because it scored below GPT-5.6 Sol on every High capability evaluation.⁵⁵

The GPT-6.1 Sol addendum later qualified Luna's safety results. Some of Luna's higher jailbreak robustness scores may reflect a broader tendency to refuse, legitimate requests included.⁸⁸8OpenAI, Addendum to GPT-6 Astra System Card: GPT-6.1 Sol, 29 September 2026, section on jailbreaks. https://deploymentsafety.openai.com/gpt-6-1-sol

On 25 September OpenAI fixed a bug in image encoding that had degraded image understanding in Luna and GPT-6 Sol. It advised rerunning any evaluation that uses image inputs.¹¹

Using it

Luna accepts reasoning effort none, which Astra and GPT-6.1 Sol reject.²² Its range runs up to max, with medium as the default, and requests use the id gpt-6-luna.³³

In Chat Completions, Luna calls functions only at effort none. Tools combined with reasoning need the Responses API.³³

Up to 272K input tokens, input costs $0.10 per million and output $0.50. Above that, the rates are $0.20 and $0.75.⁷⁷ Cached input costs $0.01, a tenth of the input price.⁷⁷

Batch and Flex cost half of Standard.³³ Fast mode is not offered with EU data residency.²²

The model selection guide puts Luna at low effort for small edits and simple data extraction.⁹⁹9OpenAI, model selection guide, read 29 September 2026. https://developers.openai.com/api/docs/guides/model-selection