In ChatGPT (web), I explicitly select “ChatGPT 5.2 Thinking” and enable “Extended Thinking”. I do this specifically to avoid any auto-routing to the Instant model. Despite that, certain prompts appear to be handled by the Instant model anyway, as if “Auto” were selected.
This is most reproducible with translation, grammar fixes, and style rewrites. It also happens intermittently when asking the assistant to explain something referenced earlier in the same conversation.
Can confirm this used to happen with 5.1 thinking. And it’s happening again with 5.2. The quality of answers is degraded noticeably when it assumes thinking is not needed. If this “bug” is not fixed, might have no choice but to cancel subscription and move to better competitors.
Did you notice this with 5.1 Thinking right after it launched? If so, did it go away after a while?
If it did, maybe it’s just an early-release quirk.
Yes it did go away. This may be temporary, but the lack of transparency about such a thing is concerning.
It stil hasn’t gone away. I think it’s permament and intended.
Yes probably permanent. I’d also wager they moved the correct behavior under the guise of “Heavy-thinking“ to the Pro plan. So, it’s another thing behind the paywall.
You still see it? This 5.2 model has been a disappointment.
2026 is about to start, and this has still not been addressed. I’m starting to think it’s intentional.
As fabiondo said, since 5.2 Thinking came out, I noticed some responses answering instantly, without the “thought for a couple of seconds” that used to appear in 5.1 Thinking. I thought it was just a UI change, but it seems it’s not.
I have a GPT (those apps you can create) meant to transcribe from Japanese to roumaji in a specific way, and I have always used it with the thinking model because the instant one always made mistakes following the instructions. Since 5.2 Thinking came out, the model went from thinking for 4-16 seconds to less than 7 seconds, sometimes responding in an instant, and making the exact same mistakes Instant makes. When I tell it it made a mistake, then it reasons for around 10 seconds and delivers an actually good response. Then if I ask for another thing in the same chat, it once again responds in an instant and makes mistakes.
We need some transparency about this. I get that OpenAI is burning money with each account, but I’d prefer them to reduce the amount of weekly thinking uses from 3000 to 250 again (like it was in the 4o days) and letting us actually use the thinking mode when we select it. Or just remove thinking mode and instant mode and being truthful about what they are selling us, 5.2 Auto. And this change came just it time to be around the holidays, when most people won’t notice.
I’m also concerned about token windows. Plus accounts get 32k context windows with Instant and 198k with Thinking. Does this routing mess with that? We’ll likely never know.
I hope more people discuss this on reddit and youtube. It’s the only way for them to actually acknowledge what’s going on, by calling them out.
The most infuriating part is the sneakiness in the way they’re doing it. They aren’t communicating it, and they’re hoping most people won’t notice. It also has an air of, ‘We know better than you, the user.’
That said, I’ve found an effective way to make it always use Thinking. Attach this to the end of the prompt: (Before you give the final output, do a deliberate internal pass, then do a brief “process summary describing what you checked or considered (no hidden reasoning, just a checklist-style summary). Do not include your process summary on the answer, it’s internal. Then give the final answer.)