Don't stop here
Hand-picked guides our readers explore right after this one.
Master ChatGPT with advanced prompting techniques, mega-prompts, and proven frameworks
Read the guideExpert guide to Claude prompts with XML tags, artifacts, and complex reasoning
Read the guideStunning image generation with Midjourney prompt mastery
Read the guideIf ChatGPT is answering complex questions instantly with shallow, generic output, the cause is almost certainly the router rather than the model being incapable. Since GPT-5, ChatGPT is a unified system: a fast instant model handles most turns and a deeper reasoning model (Thinking) is engaged only when a real-time router judges the request complex enough. The router defaults toward the fast path because that is what keeps the service affordable at scale, so a question that looks short but is actually hard gets the cheap treatment. You can override this. In the current 2026 lineup ChatGPT runs on the GPT-5.5 family (GPT-5.2 Instant, Thinking, and Pro were removed on 12 June 2026 and existing chats were migrated to the matching GPT-5.5 model), and Thinking is selectable directly in the model picker on paid plans and through the '+' menu on Free and Go with a small cap. Note that older advice telling you to switch back to GPT-4o no longer applies, since GPT-4o was retired in early 2026.
Hard questions come back instantly with no visible reasoning step
Answers are confidently wrong, generic, or skip the constraints you set
The same prompt gets a thorough answer one time and a shallow one the next
You never see a 'Thinking' or reasoning indicator before the reply
Long multi-step tasks get summarised instead of worked through
Behaviour changed after a model migration without you changing anything
The core cause. A real-time router weighs query complexity, tool needs, and context to pick between the instant model and the reasoning model, and it favours the fast path by default. Short prompts read as simple even when the underlying task is hard.
Leaving the picker on the automatic or Instant option hands the decision to the router on every turn. Selecting Thinking explicitly is what removes the guesswork.
Reasoning-model messages are metered separately, with a weekly pool on paid plans and a small per-window cap on Free and Go. Once that runs out you get routed back to the fast model instead of an error, which looks like the router ignoring you.
OpenAI retires models on a schedule (GPT-5.2 was removed on 12 June 2026, GPT-4o earlier in 2026) and moves conversations to the successor. Tuned prompts and saved workflows can behave differently afterwards even though nothing on your end changed.
A one-line question with no constraints, no data, and no requested output format gives the router nothing to latch onto. Explicitly asking for step-by-step reasoning or verification makes the deeper path much more likely.
When to try: First, and for any genuinely hard task
Open the model dropdown at the top of the chat and choose the Thinking option (GPT-5.5 Thinking on current plans) instead of Instant or Auto. On Free and Go the Thinking option sits behind the '+' menu in the composer and is capped, so use it on the turns that matter.
When to try: Every time you cannot pick the model yourself
Add an instruction like 'think this through step by step before answering, then state your assumptions and check the result'. This both nudges the router toward the deeper path and improves the fast model's output when you do get routed there.
When to try: When short prompts keep getting shallow answers
State the constraints, paste the data, name the edge cases, and say what a correct answer must satisfy. A short question about a complex problem reads as simple to the router. A detailed one does not.
When to try: Immediately after a shallow answer
If an answer is clearly under-thought, do not argue with it. Switch the picker to Thinking, then re-send the request as a fresh message rather than replying 'try again', which usually keeps the same routing.
When to try: If Thinking stops engaging even when selected
Look at your plan's limits in Settings. If the weekly Thinking pool is exhausted you will be quietly served the fast model. Wait for the window to roll over, or upgrade if you hit it every week.
When to try: If you are chasing a removed legacy model
GPT-4o was retired in early 2026 and GPT-5.2 followed on 12 June 2026, so guides telling you to switch back to a legacy model are stale. Map the task instead: Instant for drafting and quick lookups, Thinking for analysis, code, and maths, and the Pro model where your plan includes it for the hardest work.
When to try: For long or high-stakes tasks
Break a big request into a plan step, an execution step, and a verification step, each as its own message. Multi-step prompts engage deeper processing and give you a checkpoint to correct before errors compound.
When to try: After an announced model retirement
After a forced model change, re-test your custom instructions, Projects, and Custom GPT instructions against the new model and tighten anything that drifted. Prompts tuned for a retired model often need the output format restated explicitly.
Pick Thinking manually for analysis, code, and maths rather than trusting Auto
Write prompts that state constraints and success criteria, they route better
Keep an eye on your weekly Thinking allowance so it is there when you need it
Re-test tuned prompts whenever OpenAI retires a model version
Contact OpenAI via help.openai.com if the Thinking option is missing entirely from a paid account, if selecting it visibly has no effect across multiple fresh chats, or if the model picker fails to load. Routing behaviour itself is by design and is better raised as product feedback with concrete prompt examples than as a support ticket.
Because a real-time router, not you, decides between the instant model and the reasoning model on each turn, and it favours the fast path by default to keep the service viable at scale. Short prompts with no stated constraints look simple to the router even when the task is hard.
Select the Thinking option explicitly in the model picker at the top of the chat rather than leaving it on Auto or Instant. On Free and Go it is available from the '+' menu in the composer with a small cap. Adding 'think this through step by step' to the prompt reinforces it.
No. GPT-4o was retired in early 2026 and GPT-5.2 Instant, Thinking, and Pro were removed on 12 June 2026, with existing conversations moved to the matching GPT-5.5 model. Any guide telling you to switch back to a legacy model in the picker is out of date. Choose GPT-5.5 Thinking for the work you used a stronger model for.
Two likely reasons. Either your weekly Thinking allowance ran out and you are being served the fast model silently, or a scheduled model retirement migrated your conversations to a successor model that responds differently to the same prompt. Check your limits first, then re-test the prompt with Thinking selected.
Partly. It does not guarantee the reasoning model, but explicitly requesting step-by-step reasoning, stated assumptions, and a verification pass measurably improves output on either model. The reliable method is still selecting Thinking in the picker yourself.