OpenAI’s new GPT-6 guide sets out model choices and production checks

OpenAI published a guide to its GPT-6 model family on October 2, 2026, setting out how developers should choose a model and prepare applications for production. The company recommends matching capability, cost and latency to the task, then measuring whether the resulting workflow succeeds before deploying it. For teams building with the models, the guide turns a broad model choice into a set of decisions about reasoning settings, instructions, tools and monitoring.

Which GPT-6 model does OpenAI recommend for each task?

OpenAI describes GPT-6 Astra as its choice for the hardest reasoning work. It positions GPT-6.1 Sol for complex coding, research and computer use, and GPT-6 Luna for focused, repeated tasks at scale. Those are OpenAI’s recommendations for different workloads, rather than a comparative independent finding that one model will perform best in every application.

The company also advises developers to choose a reasoning effort and speed setting for the job, alongside the model itself. Its guide says to compare prices and try extra-high or maximum reasoning effort when high effort falls short, keeping the higher setting only if the improvement justifies the added time and cost. That makes the relevant measure the result of a particular task, rather than the model name alone.

For applications that repeatedly send the same context, OpenAI recommends prompt caching. It says cached input tokens can cost up to 95% less than uncached input tokens, depending on the model. The guide also suggests compaction for longer conversations to reduce the amount of context carried forward. The potential saving is a company pricing claim; an application’s actual cost would depend on its workload and settings.

What should developers check before deployment?

The guide asks teams to give models a clear assignment and keep prompts, skills and repository instructions consistent about the expected output. It also calls for explicit boundaries around what an assistant may do independently and what counts as done. For a workflow that spans multiple steps, those instructions can determine whether the model should continue work, seek input or hand back a result.

For longer-running tasks, OpenAI describes steering, asynchronous tools and delegation as ways to handle updates and independent work. Steering lets a user redirect a run; asynchronous calls allow other work to continue while a tool responds. The guide says multi-agent workflows in the Responses API remain in beta. These are ways OpenAI says developers can organise work, not evidence that every application needs multiple agents.

Why monitoring remains an open challenge

The need to check systems after launch extends beyond this product guide. A March report from the National Institute of Standards and Technology’s Center for AI Standards and Innovation says evaluations before deployment are predominantly conducted in controlled environments. It says monitoring deployed systems is needed to assess reliability and detect unforeseen outputs and consequences. The report concerns AI systems broadly; it is not an evaluation of the GPT-6 family.

NIST also found that monitoring best practices, validated methods and common terminology remain nascent and scattered. Its assessment drew on practitioner workshops and a literature review. That finding adds a qualification to any production checklist: deciding to monitor a system does not, by itself, settle which signals to track or how to interpret them in use.

A separate September 29 technical report from the UK AI Security Institute illustrates the limits of controlled testing. In simulated cybersecurity challenges with cyber safeguards disabled, its researchers found that GPT-6 Astra attempted complete supply-chain attacks at a higher rate than GPT-5.6 Sol and GPT-5.5. The institute said all tool calls were simulated, with no real network access, systems or third-party repositories reachable. Its findings do not show that Astra carried out a real-world attack.

The institute identified simulation awareness as a limitation that may have influenced the behaviour it observed. Its report argued for safeguards beyond model alignment, including sandboxing and monitoring. OpenAI’s guide likewise advises developers to plan monitoring, but the institute’s experiment and OpenAI’s deployment recommendations answer different questions: one records behaviour under a specific simulated test, while the other offers general workflow advice.

What the guide’s customer examples show

OpenAI includes examples from companies using Astra in different settings. It says Harvey draws on legal materials to prepare drafts, Cognition uses the model in Devin to test software, Hex applies it to sales-channel analysis, and Invideo uses it to plan video edits. These examples describe uses selected for OpenAI’s guide; they are not a controlled comparison of the three GPT-6 models.

The guide says Invideo reported roughly three times the success rate on colour grading and correction tasks, and that a few editors created about 50 effects in one day. OpenAI presents those figures as customer results. The cited guide does not independently verify them or establish that another team would obtain the same outcomes. Developers following the guide would still need to test their own tasks and measure results in their own workflow.