GPT-6 Luna vs Gemini 3.8 Flash: Which Cheap Model Should Carry Your Volume?
Updated 27 October 2026. OpenAI released GPT-6.1 Terra on 26 October: $1 input / $0.05 cached input / $5 output per million tokens, with a 1,050,000-token context window and 128,000 maximum output. It is a fast option for visual reasoning, computer use and natural writing. GPT-6.1 Sol and GPT-6 Astra remain better candidates for the most complex coding; Terra does not replace every frontier model. See OpenAI’s GPT-6.1 Terra announcement and our current ChatGPT model guide.
Scope: the named model comparison below retains its original models and dated examples. Terra is an additional current option, not a silent replacement for either model.
Published 23 September 2026 by Jake Hissitt, Founder of Stob.AI.
Most production AI spend is not going on hard reasoning. It goes on millions of small, repetitive calls: classify this email, extract these fields, route this record, summarise this note. For that work, two models now dominate the shortlist.
GPT-6 Luna at $0.10 in / $0.50 out, and Gemini 3.8 Flash at its introductory $0.75 in / $3.75 out through 31 December 2026. Flash moves to $1.50 in / $7.50 out from 1 January 2027.
Side by side
| GPT-6 Luna | Gemini 3.8 Flash | |
|---|---|---|
| Input / 1M | $0.10 | $0.75 through 31 Dec 2026; $1.50 from 1 Jan 2027 |
| Output / 1M | $0.50 | $3.75 through 31 Dec 2026; $7.50 from 1 Jan 2027 |
| Cached input / 1M | $0.01 | Available |
| Context | 1,050,000 | 1,048,576 |
| Max output | 128,000 | 65,536 |
| Pricing rule | Above 272K input: $0.20 / $0.75 | Introductory rates through 31 Dec 2026; standard rates from 1 Jan 2027 |
| Strengths | Low-cost tier, reasoning-effort control, image input and tool support | Native image, audio and video input; computer use, search grounding, speed |
| Published benchmarks | Shares the GPT-6 family controls and context | Terminal-Bench 2.1 89.4%, HLE-Verified 54.9%, OSWorld 2.0 59.0% |
Cost per million tasks
Assume a typical classification call: 800 input tokens, 120 output tokens, no caching. These examples use Flash's introductory rates through 31 December 2026.
| Model | Input cost | Output cost | Per 1M tasks |
|---|---|---|---|
| GPT-6 Luna | $80 | $60 | $140 |
| Gemini 3.8 Flash | $600 | $450 | $1,050 |
| GPT-6 Sol (for reference) | $1,600 | $1,200 | $2,800 |
| Claude Opus 5.5 (for reference) | $3,200 | $2,400 | $5,600 |
Luna is roughly 7.5 times cheaper than Gemini 3.8 Flash on this shape of work, and 20 times cheaper than Sol. With cached input at $0.01 per million, a Luna workflow with a stable prompt prefix gets cheaper again.
So why would you pay for Flash?
Gemini 3.8 Flash supports native audio and video input alongside images, with computer use and search grounding. Luna also accepts images and supports web search and computer-use tools through the Responses API. Compare completed outcomes on your own inputs before paying for a higher tier.
- Native audio and video input. Flash can process these media directly; Luna supports text and image input.
- Computer use. Flash reports 59.0% on OSWorld 2.0. Luna also supports computer use; benchmark the actual browser or desktop workflow.
- Search grounding. Flash integrates Google Search grounding. Luna also supports a web-search tool; compare answer quality and tool charges.
- Terminal work. Flash reports 89.4% on Terminal-Bench 2.1. This published result does not establish a direct comparison with Luna on the same setup.
Sources checked 6 November 2026: OpenAI's Luna specifications and Google's Gemini API pricing.
The rule we use
Start with Luna for structured text tasks and test its image and tool support where relevant. Evaluate Flash when native audio or video input is required, or when it produces better accepted outcomes on your workflow.
A routed system can use Luna to read and classify the inbound stream, Flash for native audio or video input and tasks where it passes stronger quality checks, and a frontier model for difficult exceptions. That layering is the practical version of the advice in 7 things before building a reliable AI system.
Frequently asked questions
What is the cheapest capable AI model in 2026?
GPT-6 Luna at $0.10 per million input tokens and $0.50 per million output is the cheapest model from a frontier family. Open-weight models can be cheaper again if you self-host.
Is Gemini 3.8 Flash worth 7 times the price of Luna?
The introductory rates are 7.5 times Luna's input and output rates through 31 December 2026. Flash may earn that premium for native audio or video input, or better outcomes on your task set. Luna also supports image input, web search and computer use, so those capabilities alone do not establish better value.
Can Luna handle long documents?
Yes, it has a 1,050,000-token context window, though prompts above 272,000 input tokens bill at $0.20 in / $0.75 out for the whole request.
Which is faster?
Both are built for low latency. Gemini 3.8 Flash is optimised for fast multimodal responses; Luna is the fastest tier in the GPT-6 family.
What to do next
Find the highest-volume model call in your stack and check what it is actually doing. If it is structured text work running on a premium model, move it to Luna and measure the accepted-outcome rate. Compare full pricing in our cheapest production AI models guide and the model comparison tool.
Related: GPT-6 tier guide · Best Gemini Model 2026