Should You Use a Thinking Setting? Compare It With Three Answerable Tests

When this article was first published, GPT-5 Thinking was a ChatGPT model name. OpenAI's retirement notice records its removal from ChatGPT in February 2026. Evaluate any current Thinking or reasoning-time option by evidence rather than transferring the old name.

Fix the comparison conditions

Use a new chat, identical input, the same files, and the same tool access for the normal and reasoning settings. Do not mix in different history or memory.

Setting:
Start and finish:
Correct answers:
Constraint violations:
Unverifiable claims:
Change on a repeated run:

Test 1: extraction

Write a ten-line source and ask for only its dates, owners, and deadlines, with no invented fields. Compare the result with your answer key.

If a simple extraction shows no improvement, there may be no reason to spend additional time on every similar task.

Test 2: constrained planning

Provide a budget, deadline, and prohibited method. Score compliance with every constraint, not how impressive the plan sounds.

Within $40 and 30 minutes, create a proofreading procedure.
Do not create an external account.
Ask rather than guess when a required assumption is missing.

Test 3: small bug fix

Use short code with a known defect and a test. Check whether the test passes and whether unrelated code changed. If execution is unavailable, score whether the limitation is stated accurately.

Decision criteria

• more correct answers or passing tests;
• fewer format or prohibited-action violations;
• fewer unsupported assertions;
• a benefit proportionate to the additional time.

OpenAI states that model output can still include errors and fabricated citations. Verify consequential facts in primary sources regardless of the reasoning setting.

Conclusion

Do not equate a Thinking label with universal accuracy. Compare extraction, constrained planning, and bug fixing under identical conditions, then record accuracy, violations, and time. Check current names in Model Release Notes and the live account interface.