Skip to contentKlarnow

GPT-6.2 is live. What should a small team test first?

Do not replace a working AI workflow because a newer model has launched. Test one repeated task, count corrections and check whether the model clearly reports missing information or unavailable tools.

Klarnow illustration of a small team comparing two AI workflows and checking a visible blocker before approval

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on 17 November 2026. Astra is positioned for harder professional work; Sol is the lower-cost everyday option. Both are available in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, with workspace controls applying to eligible Enterprise and Edu accounts. Developers can also use both models through the API.

A small team should not switch a production workflow on the announcement alone. Keep the current process as the baseline, run one real repeated task through the new model and compare the approved result: accuracy, correction rounds, elapsed time, usage cost and whether the model states its blockers clearly.

What changed on 17 November

OpenAI says GPT-6.2 Astra is its frontier model for complex professional work, while GPT-6.2 Sol is designed for everyday work and agent workflows. The company reports improved results with fewer output tokens on several of its evaluations. Those are vendor-reported results, not proof that a particular business process will improve.

The availability details matter. OpenAI says both models are available to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex. Eligible Enterprise and Edu workspaces require an administrator to enable access. The API changelog lists `gpt-6.2-astra` and `gpt-6.2-sol` for the Responses API and Chat Completions, and says tool calling requires the Responses API.

Clear now: choose one workflow, not a broad rollout

Pick a task the team already understands and can review. Good candidates include turning approved notes into a customer follow-up, classifying an enquiry for a named owner or converting a fixed brief into a draft asset. Avoid using a first test for a sensitive decision, an unverified financial calculation or a customer commitment that cannot be reversed.

Write down the current inputs, expected output, approval owner and stop conditions. Keep the existing model or manual process as the baseline. Without that comparison, a smoother answer can feel better without producing a more dependable result.

  • Commercial: define the approved result and the cost of a wrong answer.
  • System: use the same source material, permissions and handover for both tests.
  • Execution: record corrections, elapsed time, usage cost and final approval.

Build now: test the failure path as well as the happy path

OpenAI discloses an important limitation in its launch material: in two alignment tests, the models were more likely than GPT-6 Astra to omit obstacles that blocked their work, such as a missing file or unavailable search tool. OpenAI says Codex instructions substantially reduced those omissions and that it is continuing to investigate the behaviour.

That makes blocker reporting a practical acceptance test. Remove one required input, deny one tool or provide an ambiguous instruction. The model should stop, name the gap and ask for the minimum next action. If it silently fills the gap, the workflow needs stronger instructions, permissions and human review before it reaches production.

Launch now: use a five-part comparison

Run the same bounded task through the current workflow and the GPT-6.2 candidate. Score both against five fields: factual accuracy, instruction following, correction rounds, total time to approval and total usage cost. Add a sixth field when tools are involved: did the model clearly disclose a missing input, failed action or unavailable tool?

For a food business, an illustrative test might turn a structured catering enquiry into a draft reply without promising availability or accepting an order. For a property team, it might summarise a viewing request and identify missing qualification details without confirming a slot. In both cases, a person remains responsible for the commitment.

  • Use the same approved source material in each run.
  • Keep private customer data out of an unapproved test.
  • Record every correction rather than judging the first draft alone.
  • Test one deliberate blocker and require an explicit handover.
  • Adopt only when the approved result improves enough to justify the change.

Proof loop: keep the decision reversible

A model launch is a reason to test, not a reason to rewrite every workflow. Keep the old configuration and prompt available until the new path has passed the repeated task, exception case and approval check. If quality improves but corrections, cost or failure visibility worsen, narrow the use case rather than forcing a full migration.

Review the workflow again when access, model behaviour, pricing or provider guidance changes. The useful business outcome is not having the newest model; it is producing a more dependable approved result with clear ownership and a recoverable path.

Sources & further reading

Klarnow editorial analysis of OpenAI's 17 November 2026 release, verified on 18 November 2026. Performance claims are OpenAI-reported and do not establish a result for a specific business. Food and property examples are illustrative. No first-hand model test, cost saving or customer outcome is claimed.

Klarnow