gpt-6-luna

Overview. GPT-6 Luna is OpenAI's fast, cost-efficient model in the GPT-6 family, released in September 2026 as the lightweight counterpart to the flagship GPT-6 Sol reasoning model. It combines a 1,050,000-token context window with a 128,000-token output ceiling and six dialable reasoning-effort levels, from none to max. Developers reach for it when they need high-volume chat, classification, or lightweight agentic responses without the overhead of a full reasoning model.

Key strengths. Luna's defining feature is its reasoning-effort dial: a single model can serve instant, low-effort replies or step up to deeper multi-step responses depending on the request, without switching endpoints. It supports text-and-image input, tool calling, structured output, and context caching, which together suit latency-sensitive agentic pipelines and high-throughput chat or classification workloads. OpenAI positions it explicitly for chat, classification, and lightweight agentic tasks rather than deep specialist reasoning.

When to use it. Use Luna for realtime chat, classification, and lightweight agent loops where response speed matters more than deep multi-step analysis. Its large context window makes it a practical choice for ingesting long documents or extended conversation histories, and the reasoning-effort control lets you selectively spend more compute only on the requests that need it. It's a weaker fit for tasks that require deep, multi-step specialist reasoning — those are better served by its GPT-6 Sol sibling, which OpenAI positions as the flagship reasoning model.

Limitations. OpenAI has not published parameter counts, architecture, tokenizer, or Mixture-of-Experts details for Luna, and the model is closed-weight with no self-hosting path. Its training data cutoff is May 18, 2026, and it is tuned for breadth and speed over the deeper reasoning depth of its GPT-6 Sol sibling.

Sovereignty note. GPT-6 Luna is built by OpenAI, a US-based creator, so deployments carry standard US CLOUD Act exposure. However, OpenAI's own documentation states that EU data residency is available for this model when using Standard processing. Training-on-input policy specifics for this model were not confirmed in research and should be checked directly with OpenAI before sensitive deployments.