The GPT-6 family raises a practical selection question, and OpenAI answered it systematically in its official guide, A model guide for the GPT-6 family, published on October 2, 2026. The guide does more than list prices for three tiers. It covers reasoning levels, prompt caching, prompting, and long-task management together, giving development teams a method for configuring cost and capability as one decision.
Three tiers and their prices
| Model | Input | Output | Cached input | Positioning |
|---|---|---|---|---|
| GPT-6 Astra | $10 | $50 | $1 | The hardest reasoning work, when the highest intelligence is needed |
| GPT-6.1 Sol | $2 | $10 | $0.10 | Intelligence close to Astra at one-fifth the price |
| GPT-6 Luna | $0.10 | $0.50 | $0.01 | Well-defined, large-scale, repetitive everyday work |
All prices are per million tokens. GPT-6.1 Sol suits complex programming, research, and computer use. GPT-6 Luna is aimed at repeatable tasks such as extracting invoice fields, classifying requests, and generating structured summaries. With gaps of up to a hundredfold between tiers, choosing the wrong model multiplies cost directly.
Choose the model and the reasoning level together
The core principle is to treat model choice and reasoning level together as a trade-off between intelligence and price. Low suits routine work such as extracting facts and making small edits. Medium suits work that requires judgment, such as planning a feature or comparing options. High suits difficult debugging, deep analysis, and careful review. Extra high and Max should only be tested when High is not enough, and kept only when the improvement is worth the extra time and cost. Teams should not default to the top setting for everything; they should schedule by task difficulty.
Prompt caching and context management
Cached input can cost up to 95% less than uncached input. The guide recommends placing stable instructions and reference material before the changing task details and keeping tool definitions consistent to improve cache hits. For long conversations, compaction can compress the context while preserving the state needed to continue the work, keeping long tasks from growing ever more expensive and slow.
Prompting is changing
Eric Provencher, OpenAI's head of developer experience, says models have become better at understanding nuance and ambiguity, and that overly specific guidance that used to help can now get in the way of results. The guide asks users to give the model a clear brief: the desired result, who it is for, the relevant background and constraints, and what counts as done. Descriptions of skills should be short and clear, stating when they should run. Explicit decision boundaries should replace blanket "ask me first" rules. And "done" means implementing the change, running it, checking the result, and fixing failures.
Designed for tasks lasting hours to days
The GPT-6 family can take on tasks lasting from hours to days. On the API side, the Responses WebSocket API lets users steer the model mid-course while it works; updates are queued and do not cancel a tool already running. While a slow tool runs, the model can first do work that does not depend on it. GPT-6.1 Sol can also assign independent subtasks to sub-agents in parallel in the Responses API, with multi-agent support currently in beta. All three models can operate websites and desktop applications directly (computer use), with one principle: choose the most reliable and simplest method at each step, and when a task can be completed through an API or a connected tool, do not read the screen and click buttons instead.
Team cases provide reference points. Cognition used Astra to trace and verify, end to end and with no prior information, a fix for a payment bug involving $5,000 in cash and 200,000 in API credits, within 30 minutes. Harvey used Astra with audit logs to test a legal agent over a corpus of nearly 8,000 documents. Invideo used Astra for color grading and correction tasks and roughly tripled its success rate.
One caveat: the guide does not publish the context window sizes of the three models, so teams planning very long-context scenarios will have to wait for further information. For teams choosing now, the guide's value is that it turns the decision from "pick the strongest" into "pick by task layer": Luna to keep everyday costs down, Sol for complex work, Astra for genuinely hard problems, with reasoning levels and caching layered on top so overall spending stays under control.
