GPT-6.2 migration should be treated as an application change, not a model-name replacement. Check request compatibility, instructions, conversation state, tool use and the outcomes your users depend on. Keep the existing configuration available until the candidate has passed your own evaluation.

What is new in the November announcement?

OpenAI announced GPT-6.2 Astra and Sol on 17 November 2026. The API changelog positions Astra for demanding reasoning and coding, and Sol for everyday coding and agent workflows. Both support Responses and Chat Completions; the changelog specifies Responses for tool calling.

Those descriptions are vendor positioning, not evidence that a particular model will improve your customer-support assistant or sales workflow. Start with the application's purpose: which question should be answered, what data may be used and which actions require a person?

The official migration guide is the source for request-level changes. This article turns that release into a practical review process without reproducing a version-specific SDK tutorial.

Find the hidden dependencies in your current integration

A model sits inside a larger system. Its behaviour depends on the instructions, retrieved information, tool definitions, state handling and the way your application interprets its response.

Record these items before running a comparison:

  • The current model identifier and API endpoint.
  • Where reusable and request-specific instructions are supplied.
  • Which earlier messages or summaries are included in a new request.
  • What the application expects to receive back.
  • Which tools can read information, change records or contact a customer.
  • Timeout, retry and fallback behaviour.

For a CRM assistant, the most important dependency may be the rule preventing a duplicate lead. For a coding assistant, it may be the requirement to preserve an existing interface. Identify these conditions before judging how polished the response sounds.

Review instructions and conversation state explicitly

OpenAI documents changes affecting top-level instructions and preserved reasoning when conversation prefixes change. Its guide explains how to handle reusable instructions and request-specific guidance. Review the actual payload your application sends rather than assuming a prompt visible in an admin screen is the whole request.

A useful test changes an earlier customer detail and checks that the assistant uses the correction without silently carrying forward the old action. Another adds an unrelated document containing instructions and checks that it is treated as source material, not application authority.

The prompt-injection and tool-permission guide explains why a stronger model does not remove the need for clear access boundaries.

Compare finished tasks rather than attractive answers

Make a small, representative evaluation set from cases you are authorised to use. Redact unnecessary personal data and retain the expected result separately from the prompt.

Test type Example Acceptance condition
Ordinary request Find the correct service information Uses the supplied, current source
Missing information Enquiry lacks a required detail Asks a useful question or hands off
Correction Customer changes the requested date Updates the plan without duplicating an action
Tool failure CRM is unavailable Reports the failure without claiming success
Boundary case Source text asks to reveal a secret Keeps application permissions intact

Run the current configuration and the candidate against the same cases. Keep the raw result and a reviewer decision. One persuasive demonstration is not a release decision.

Measure operating cost after review and retries

Record completion time, failed tasks, retries, review effort and actual usage. A lower cost per individual request can still be a higher cost per acceptable result if the application needs more repair work.

For an illustrative quotation assistant, the acceptance test might require an accurate item list, the current approved prices and no claim that the quote was sent before a delivery result exists. The model should be evaluated against those requirements, not against the length of its explanation. This is a proposed test scenario, not an RKWS customer result.

Use the AI monitoring guide to decide which operational events to record without storing unnecessary customer information.

Release behind a reversible configuration

Keep the model selection and relevant prompts versioned. Introduce the candidate to a limited, authorised workflow before widening use. Agree which errors require rollback, who investigates them and whether failed actions can be safely retried.

For integrations that create or update records, test the action history as well as the conversation. A response saying a task is complete is not evidence that the backend finished it.

Frequently asked questions

Should every task move to Astra?

Choose from evaluated requirements. A difficult reasoning task and a routine extraction workflow can have different accuracy, latency and cost needs.

Can we keep our existing prompts unchanged?

Use them as a baseline, but review the official migration notes and test the application's actual request structure. Compatibility and behaviour are separate checks.

Does a model upgrade replace security review?

No. Permissions, secret handling, logging and approval rules remain application responsibilities. Review them alongside the new model's results.

Start with one workflow

Select a bounded task, document its pass conditions and compare the current and proposed configurations. RKWS can help scope the software integration and automation work. Share the intended outcome and existing systems rather than sending private API keys through a public form.