Business workflows
WinnerAutomationBench 1.0.6
At high effort, Luna gains 5.4 points over its predecessor. This is OpenAI’s 47-tool workflow comparison, separate from AutomationBench-AA.
SPIRITT
GPT-6 LunaGPT-6 Luna makes scale affordable
Extract data, triage queues, classify documents, and run focused agents without giving every request a flagship-sized bill. GPT-6 Luna brings a 1.05M context window at $0.10 input and $0.50 output per million tokens.
GPT-6 Luna in action
Explore low-cost websites, 3D scenes, CAD, and creative workflows, alongside firsthand notes on where Luna shines and where it needs more help.
A lighter bill for work that repeats
Use Luna where the tasks are focused and the volume adds up. Save the heavier model budget for the decisions and projects that need it.
Focused tasks, much more breathing room
OpenAI positions GPT-6 Luna as its most efficient model for focused, high-volume work. It supports text and image input, up to 922K input tokens within a 1.05M context window, and up to 128K output tokens. Medium is the default effort; harder cases can use higher settings.
Standard API prices are $0.10/M input, $0.50/M output, and $0.01/M cached input. Compared with GPT-5.6 Luna, input is 50% cheaper and output is about 58% cheaper. Both standard rates are one twentieth of GPT-6 Sol’s.
Artificial Analysis reports roughly $0.07 per Intelligence Index task at max effort, versus $0.18 for GPT-5.6 Luna. Its Coding Agent Index falls from 43 to 41, so the savings are not a promise of better results on every task. Use Luna for focused work and check that the output is complete.
Focused agentsHigh-volume work1.05M context128K output$0.10 / $0.50 per M$0.01/M cached input
AA Intelligence Index v4.3
Artificial Analysis, September 22. OpenAI family comparison at max effort. Luna and its predecessor both round to 37, at different costs.
Where Luna makes the budget go further
See the balance between coding, workflows, computer use, and cost. Lower prices are the headline; capability and completeness still depend on the task.
Workflow economics
WinnerAutomationBench cost (US cents/task)
Same high-effort settings as above: 2.08 versus 5 cents per task, about 58% lower. Source: OpenAI.
Long-running work
Agents’ Last Exam V1
Long-horizon professional workflows. Opus 5 uses high, its best evaluated setting in this source; the OpenAI rows use max. Source: OpenAI.
Merge-ready code
FrontierCode 1.1 Main
Changes are judged for correctness and merge-readiness. Fable 5.1 medium is shown rather than a weaker higher-effort run. Source: OpenAI.
Software engineering
DeepSWE v1.1: OpenAI evaluation
Complex repository work. Efforts are labeled; these official research/API runs differ from Artificial Analysis’s Codex evaluation below.
Computer use
OSWorld 2.0 offline
Partial reward on the v2026.08.08 offline set, not full-task completion. Both generations are shown at max; any other effort is labeled. Source: OpenAI.
Factual reliability
Factual error rate on difficult prompts
Lower is better. Prompts were selected from user-flagged mistakes, so these are not ordinary-usage error rates. Max effort. Source: OpenAI.
Independent coding
AA Coding Agent Index v1.5
Artificial Analysis, September 22. Codex for OpenAI; Claude Code for Claude, with fallback for Fable. All rows use max effort.
Independent repository work
DeepSWE v1.1 in Codex
Artificial Analysis’s Coding Agent evaluation at max effort. Same Codex setup for both generations; separate from the official OpenAI comparison.
Independent code understanding
SWE-Atlas-QnA in Codex
Repository questions in the same Codex evaluation at max effort. Published in Artificial Analysis’s September 22 launch report.
Independent cost comparison
WinnerAA Intelligence Index cost per task
September 22 v4.3 launch snapshot, max effort. The models use more output tokens than their predecessors; the savings are driven by lower prices.
Knowledge reliability
WinnerAA-Omniscience hallucination rate
Lower is better. AA’s special hallucination metric is not an all-response error rate. Accuracy is shown separately; declining questions can change both.
Knowledge trade-off
WinnerAA-Omniscience accuracy
Correct answers across all questions at max effort. Luna’s accuracy is broadly level while its hallucination metric improves. Source: Artificial Analysis.
Communication and trust
Coding-deception rate
Lower is better. OpenAI deliberately selected difficult, dishonesty-inducing coding tasks. Max effort; not a typical usage rate.
Tool-use transparency
Failure to disclose broken search
Lower is better. This challenge tests whether an agent admits that its search tool is broken. Max effort; not a typical usage rate.
Published API pricing
WinnerInput price per million tokens
OpenAI Standard rates for requests up to 272K input tokens. Both generations’ input price drops by 50%.
Published API pricing
WinnerOutput price per million tokens
OpenAI Standard token rates. Luna’s output falls from $1.20 to $0.50, about 58.3% lower rather than exactly half.
Published API pricing
WinnerCache-read price per million tokens
Cached reads cost 10% of uncached input. Repeated project context can use the discount when cache requirements are met.
Independent output speed
WinnerOutput tokens per second
Artificial Analysis snapshot, September 23, OpenAI API at max effort. Speed after the first chunk, not a guarantee of total task time.
Sources: OpenAI’s September 22 launch and GPT-6 system-card appendix, official API documentation, and Artificial Analysis. Each chart keeps its evaluator, version, effort, and fallback conditions. OpenAI’s comparator runs come from public reports; Fable 5 appears where 5.1 results were unavailable. The AA Intelligence and Coding Agent launch snapshots use v4.3 and v1.5 respectively. Prices are OpenAI API list rates, not SPIRITT plan prices; long-context, tool, and optional service charges can apply. Output speed is a separate September 23 observation.
How It Works
Give the work a home, choose the right model, and keep the result moving.
01
Bring the work into SPIRITT
Start with your files, repository, or business goal. A workspace brings your agent, browser, terminal, integrations, and memory together around the project.
02
Match the model to the job
Check the models and connection options enabled in your workspace. Choose the level of capability and cost that fits the work, then set a clear finish line.
03
Check the result, then keep going
Work toward a usable output: a tested change, a finished report, or a repeatable workflow. Keep the files and decisions together, and review important actions before they go live.
Put your next project to work
Bring your agent the tools, context, and workspace to turn a request into something useful.
Questions
Frequently asked questions
01What is GPT-6 Luna best suited for?+
Focused, high-volume work such as classification, extraction, triage, and smaller agent tasks. Give it clear instructions and a checkable result. Use a stronger model for the cases where errors or incomplete work would outweigh the savings.
02Is Luna exactly 50% cheaper?+
Input pricing falls from $0.20 to $0.10 per million tokens, a 50% reduction. Output falls from $1.20 to $0.50, about 58.3% lower. Cached input is $0.01/M and cache writes are $0.125/M. Actual task savings vary with tokens, effort, tools, and processing mode. These are OpenAI API list prices, not SPIRITT plan prices.
03Is Luna as capable as Sol because the context window is the same?+
No. Both support 1.05M context, but that is the amount of material they can work with, not a promise of equal reasoning quality. Sol is designed for more complex coding and agent work; Luna prioritizes efficiency on focused tasks.
04Where do the independent results show trade-offs?+
Artificial Analysis reports a lower cost per task and fewer hallucinations, but its Coding Agent Index falls from 43 to 41 compared with GPT-5.6 Luna. It also observes regressions in GDPval-AA and AA-Briefcase, including deliverables that leave out required elements. Check that outputs are complete, not just inexpensive.
05How much context can it use?+
The documented total context window is 1,050,000 tokens, with a maximum of 922,000 input tokens and 128,000 output tokens. A large window helps keep more material together; it does not guarantee that every detail will be used correctly.
06Do long prompts cost extra?+
Yes. Above 272K input tokens, the entire request is billed at 2x the normal input and cache rates and 1.5x the output rate. Tool calls and optional processing settings can add charges too.
07What do Batch, Flex, and Fast mode change?+
Batch and Flex use 50% of Standard token rates, with different latency and availability trade-offs. Fast mode costs 2x the applicable rates in exchange for higher throughput; it is not a guarantee of a particular total task speed. Regional processing adds 10% where available, and EU data residency requires Standard processing.
08Can it create images or use the computer?+
Its native inputs are text and images, and its native output is text. Through the Responses API it can use tools for image generation, web and file search, code execution, hosted shell, computer use, and more. Visual demos use those tools; they are not evidence of native audio or video output.
09Can I use GPT-6 Luna in my SPIRITT Workspace?+
Check the models and connected options enabled in your workspace. Access to a newly released model depends on the available connection and workspace setup; its OpenAI release does not automatically enable it in every workspace.
Buy from builders who use what they sellBuilt using