GPT-6 Luna cost model

Updated September 23, 2026 · first published September 23, 2026

GPT-6 Luna is the cheap model in OpenAI's GPT-6 family, released on 22 September 2026 alongside GPT-6 Sol. OpenAI aims it at high-volume work with a clear goal: summarising, extraction and quick questions. At $0.10 per 1M input tokens it is the model to route to first, and to move off only when the task proves it needs more.

GPT-6 Luna prices

TierInputCached inputOutput
Standard$0.10$0.01$0.50
Batch or flex$0.05$0.005$0.25
Fast mode$0.20$0.02$1.00
Long context (over 272K input)$0.20$0.02$0.75

Per 1M tokens, from OpenAI's pricing page. A cache write costs $0.125. Luna has the same 1,050,000-token context window as the rest of GPT-6, a May 2026 knowledge cutoff (the newest of the three) and the full effort range from none to max.

Same score as GPT-5.6 Luna, 61% cheaper per task

Artificial Analysis ran both Luna generations through its Intelligence Index (v4.3.2) at max effort.

ModelIndex scoreCost per taskOutput speed
GPT-6 Luna (max)37$0.07132.2 tokens/s
GPT-5.6 Luna (max)37$0.18142.3 tokens/s
GPT-6 Sol (low)33.9$0.13

The score did not move. The cost per task fell 61%, mostly from the lower per-token price. GPT-6 Luna at max also beats GPT-6 Sol at low effort on both score and cost. If you were running Sol at low effort to save money, Luna at max is the better deal on this index.

What high volume costs

One million classification requests a month, each with 600 input tokens and 100 output tokens, no caching:

ModelMonthly cost
GPT-6 Luna, batch$55
GPT-6 Luna$110
GPT-5.6 Luna$240
Claude Haiku 4.5$1,100
GPT-6 Sol$2,200
GPT-6 Astra$11,000

Claude Haiku 4.5 costs 10× more per token and scored 15 on the same index, as a non-reasoning model. Moving the same work from Sol to Luna cuts the bill by 95%, which is why a router that sends only the hard cases to Sol is worth building.

The catch: latency and verbosity

The index numbers above are at max effort, and at max effort Luna is slow to start. Artificial Analysis measured 104 seconds to first token. It also used 150M output tokens across the index, against a median of 84M, which Artificial Analysis describes as somewhat verbose. That verbosity is already inside the $0.07 per task, so the cost figure holds, but the wait does not suit a user-facing chat.

For interactive traffic, run Luna at none, low or the default medium and measure accuracy on your own prompts. The scores on this page are for max effort only, so accuracy at lower settings is something you have to test, not look up. Keep max effort for batch jobs, where latency does not matter and batch pricing halves the bill again.

Rate limits scale further than Sol

Luna starts at 500K tokens per minute on Tier 1, the same as Sol, but reaches 180M tokens per minute at Tier 5 (Sol stops at 40M), with up to 30,000 requests per minute. For bulk pipelines, that headroom matters as much as the price.

How to use Luna

Sources: OpenAI GPT-6 Luna model page, Artificial Analysis: GPT-6 Luna, Artificial Analysis: GPT-5.6 Luna, TechCrunch. Prices read 23 September 2026.

Related


Want this applied to your own LLM spend? FinOps LLM runs a free audit of your AI costs and shows where the savings are. Book free audit →

Back to research