Why is ChatGPT so slow?
- 01You are on a higher reasoning level than you think. ChatGPT's picker exposes reasoning effort directly — Instant, Medium, High, Extra High, plus the Pro options. These are not quality grades that happen to differ in speed; the slower ones are slower because they think for longer, and OpenAI tunes that time deliberately as a "balance between answer quality and response speed". A conversation that started on Instant and now takes 40 seconds a turn is usually a conversation that changed level.PICKER · REASONING
- 02Instant escalated to Medium without telling you. This is the cause almost nobody knows about, and it is documented in one line of OpenAI's release notes: users "have the ability to decide whether Instant auto-switches to Medium for higher reasoning when required". Auto-switching is a setting — which means it is on for people who never chose it. Your fast model is quietly becoming a slower one whenever it judges the question hard enough.AUTO-SWITCH
- 03You hit a rate limit and got moved to a fallback model. OpenAI runs smaller "Instant Mini" models specifically as "the fallback model users reach after hitting their rate limits", and notes they do not appear in the model picker. There is no banner. The session simply continues on a different model, which can feel slower, shallower or both — and no amount of restarting the app changes it, because the cause is your allowance rather than your device.RATE LIMIT
- 04It is genuinely busy, and you are not paying for the queue. OpenAI lists "priority access during high-traffic periods" as a paid benefit and describes usage limits tightening "especially during high demand". Free-tier slowness that clusters at predictable hours is this. It is also the only cause on this list you cannot fix with a setting — you either wait or you change tier.PEAK · DEMAND
Open Settings → General and turn off Instant auto-switching, then set the picker to Instant explicitly. That removes causes 01 and 02 in one move. If it is still slow, check status.openai.com before touching anything else — an incident looks exactly like a slow model.
Three of the four causes are settings, not outages.
We did not measure response times, and this page publishes no latency figures — a stopwatch on our connection would say nothing about yours. What it does instead is map the causes that OpenAI documents, because in this product the interesting slowness is configured rather than accidental.
The model picker is the centre of it. OpenAI's current picker exposes reasoning effort as the choice — Instant, Medium, High, Extra High, with Pro Standard and Pro Extended on Pro plans — and the release note explaining it describes the trade in plain terms. Thinking time is also adjusted by OpenAI over time: their own changelog records lowering Standard thinking time in January 2026, reducing it again in February, and separately restoring an Extended setting that had been changed by mistake. Speed here is a dial, and both you and the vendor have a hand on it.
The auto-switch line matters more than its length suggests. If Instant can escalate to Medium "when required", then the fast option is not reliably fast, and the user who complains that ChatGPT "randomly" takes twenty seconds is describing exactly the intended behaviour. It is switchable in General settings.
The rate-limit fallback is the one with no visible signal at all. Because the fallback model is deliberately absent from the picker, nothing in the interface tells you that you are no longer talking to the model you selected. If slowness arrives together with noticeably flatter answers, and especially late in a heavy session, that is the shape of this cause rather than of a network problem.
The causes people guess, and why they're usually wrong.
Occasionally true, and worth ruling out in thirty seconds: try a private window with extensions disabled, or the mobile app on cellular data. If it is fast there, the problem is local.
But the reason this advice dominates search results is that it is the only advice a generic troubleshooting article can give. The specific causes — reasoning level, auto-escalation, fallback models — are invisible to anyone who has not read OpenAI's release notes, so the content that ranks defaults to cache-clearing. Clearing your cache does not change which model is answering you.
Partly true, and the honest version is more useful. Thinking time is not comparable across models — OpenAI states each model is tuned independently — so a level called "Standard" on one model is not the same amount of thinking as on another. When a new model ships, the same picker setting can mean a different wait.
What is not true is that a newer model is uniformly slower. OpenAI also ships in the other direction: Fast answers returns high-confidence responses to common information-seeking questions more quickly, is on by default across plans for logged-in and logged-out users, and can be turned off in Personalization settings if you would rather always get a personalized answer. Faster and slower are both being tuned at once.
Not every wait is a fault. OpenAI's own documentation describes image generation as potentially taking a few minutes depending on complexity, and notes you can keep using ChatGPT while it works. Deep research is minutes by design. Agentic work that browses, opens files and runs tools is bounded by those steps, not by token generation.
The distinction worth holding: a chat reply that takes 40 seconds is usually a configuration story; a research task that takes 6 minutes is usually the product working. If you are paying for a tier to make the second kind faster, you are buying the wrong thing — that is a job for a bigger allowance, not more speed.
There is a failure mode that presents as extreme slowness and is not slowness at all — a request that never returns, a spinner that outlives the tab, an app that reconnects in a loop. That is the territory of our page on why ChatGPT is not working, and the first move there is different: check the status page before changing any setting.
The rule of thumb we would use: if answers arrive but late, work through the four causes above. If answers do not arrive at all, stop tuning and go and look at whether the service is up.
Where the fast answers live.
Three routes when latency is the thing you are optimising for. We publish no comparative latency benchmarks — we have not run them, and vendor-reported speed is not a measurement we would repeat as our own.
Frequently asked.
Quick follow-ups people search after this question.