OpenAI details deployment practices for GPT-6 models
OpenAI has published a practical guide for deploying its GPT-6 model family: GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna. The company positions Astra for the hardest reasoning tasks, Sol for complex coding, research and computer use, and Luna for focused, repeatable work at scale such as invoice-field extraction, request classification and structured summaries.
The guide focuses on production choices rather than a single default configuration. OpenAI advises teams to balance capability, price and latency by selecting a model, reasoning level and processing speed that match the task. It also calls for representative testing before deployment, with task success, latency and cost per successful task measured as core operational metrics.
Context efficiency and production controls
For recurring workflows, OpenAI recommends prompt caching and says cached input tokens can cost up to 95% less than uncached input tokens, depending on the model. Stable instructions and reference material should precede changing task details, while tool definitions should remain consistent so shared context can be reused. Teams should include cache writes and long-context rates when calculating the full cost of a workflow.
Longer conversations can use compaction to reduce context size while retaining the state needed to continue. OpenAI also advises removing context that a task does not need, running independent tasks together where possible, and deciding how model behaviour will be monitored. Data controls should be reviewed before an application is put into production.
Model effort, speed and instructions
In the API, reasoning effort can be set from low for routine fact extraction or small edits to high for difficult debugging, deeper analysis and careful review. Extra high or Max should be tested only where High does not deliver sufficient results, OpenAI says, and retained only when the improvement justifies the added time and cost. Reasoning effort can be changed mid-conversation without breaking cache.
Fast mode is intended for use cases where response time matters, including chat applications and coding tools, at a higher per-token cost than Standard processing. Ultrafast is available for GPT-6 Astra in Codex and the API when faster token generation is worth the premium. It operates independently of reasoning effort.
OpenAI also urges developers to state the desired result, intended audience, relevant context, constraints and definition of done. Instructions should identify actions that can proceed independently, actions that need approval, relevant documents and tests, and the required handoff. The guidance cautions that overly specific directions can hinder results when models can already understand nuance and ambiguity.
Managing work that lasts hours or days
The API supports mid-turn steering through the Responses WebSocket API, allowing corrections while a model works. Those updates are queued and do not cancel active tools or reverse completed actions. Asynchronous tool calling allows a model to continue independent work while an application runs a slower task, such as a test, before returning the tool result.
GPT-6.1 Sol also supports multi-agent workflows in the Responses API, currently in beta. It can delegate independent subtasks, such as investigating separate areas of a codebase, and combine findings into a final response. In Codex, GPT-6 Astra can request clarification during work, while users can steer an active task when requirements change.
Computer use is available in GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna for interaction with websites and desktop applications, including systems without an API. OpenAI recommends direct APIs or connected tools where they can complete a step reliably, reserving computer use for screen reading, clicking and form filling. For businesses, the implication is to begin with representative workflow benchmarks, explicit decision boundaries and monitored cost-per-success before expanding an agent into more autonomous production work.