ChatGPT has been silently routing Pro to mini for several months OpenAI's $200-per-month ChatGPT Pro plan has been silently routing some Pro model sessions to the mini model for several months, according to a user who analyzed chat metadata and contacted support for two months without a resolution. The affected sessions showed no visual cue, and the model identified itself as GPT-5 mini, with metadata properties like 'i-cot' and 'gpt-5-5-mini' indicating the downgrade. The user publicly asks OpenAI to explain the issue, citing a lack of transparency for paying subscribers. ChatGPT has been quietly downgrading Pro model into mini. No response. TL;DR Using $200/month OpenAI's ChatGPT Pro not lite , ChatGPT has an issue where some Pro model sessions responded like a mini model, or were routed into one, even when Pro was selected in model picker. This is not normal, never been documented, and has no visual cue whether it's affected or not. My finding relies on analyzing "frozen/static" chat metadata content resolved model slug rather than potentially hallucinated model outputs or chat responses. I contacted support for two months img1 fnimg1 but received only general/simple responses. Users paying for the $200 plan expect research-grade intelligence, not a same model offered to unpaid users. I publicly ask OpenAI to explain this issue. This happened next day after I subscribed to the $200/month ChatGPT Pro plan 1 fn1 . While working as usual, the responses felt different. I can understand OpenAI's model performance can be inconsistent, but this change was too extreme to ignore. The model's outputs fell far short of an expert or researcher quality, with the reasoning and thinking time significantly faster than normal by dozens of times . Consequently, the performance was significantly below what is typically expected from the Pro model. After noticing this, I asked, "What model are you?" It identified itself as a completely different model: "GPT-5 mini." At first, I thought this was a model hallucination. Since there was no technical report for the Pro model, I thought its internal system might use mini models in some cases. However, my suspicion increased after comparing different clients logged into the same account. Some clients that had remained logged in for a long time still not exact, some clients were affected too still behaved as expected from the Pro model. New logins and other clients showed invalid behavior. This must not simply a response varied with prompt complexity like simple prompt goes to a simpler model . Both the web and Android clients were affected just in case. The affected Pro sessions showed the following behavior: - The Thinking UI did not appear, or reasoning finished almost instantly. I never saw text such as "Thought for ...s" in the suspected chats. - The model always identified itself as mini when asked. - I haven't reached any Pro model usage limit. When the Pro quota is exhausted, the ChatGPT UI disables the Pro option. It did not happen here at that moment. I then investigated ChatGPT data exports and backend network responses. Based on an early reply from customer support, every chat in which the assistant identified itself as GPT-5 mini contained i-cot in the resolved model slug metadata property. About a month later, another colleague researcher reported a similar case. In that case, that property was set to gpt-5-5-mini . While investigating that report, I also found a growing number of similar issue reports on social media recently reports fnreports . Thus I decided to write about this issue. Customer support made this issue even more frustrating. I contacted support for about two months img1 fnimg1 . The replies focused on general explanations, such as model hallucinations or clearing the browser cache. I never received any explainer, actual investigation results, or clear solution. Attempts to contact relevant OpenAI employees through X also received no response img2 fnimg2 . What I consider most important about this issue is transparency . People pay $200 per month not only for Codex, but also for research-grade reasoning. Yet this behavior continued without any clear notice in the UI or by email. There was no visible indication/cue that something had changed. The issue might not be limited to Pro users; Pro Lite and Business accounts have fewer quotas. Users have no reliable way to confirm whether they are receiving the service they paid for. I was affected for several days. For that period, the money was effectively wasted away. AI products are no longer toys nowadays. They are becoming essential tools for research, development, writing, analysis, and decision-making. Users of expensive services expect more than better answers; they also expect predictable service like usual. When a model's behavior suddenly changes and performance becomes worse, the issue is no longer just performance. It becomes a matter of trust . I am not claiming that these observations prove every part of my point. However, OpenAI has provided no explanation at all. Users have every right to question whether they are receiving the service they paid for. When a provider stays silent or repeats useless answers, users lose trust in that provider. I'm not asking OpenAI to disclose confidential internal information. If this was a bug, respond that it was a bug. If it was part of an abuse-prevention or rate-limit policy, explain that effectively. If fallback routing exists, explain when it occurs, address how it appears in the UI, and explain what level of service is guaranteed to paying users. Customer support never gave me any information about any of this. Since this article is solely focused for scientists/researchers/developers , I still have a lot of stories to talk about. So I might publish more articles when they do not address this. Thank you to everyone who read this long and low effort post. About Me Computer engineering undergraduate student in South Korea. I have some LLM-like knowledge through discussions with NLP doctoral student, AI mostly vision, but work for all domains researcher who studied neuroscience on their own, and industry professionals. I do research on computer graphics, desktop Linux, and software engineering. How to Check Whether You Were Affected The following ChatGPT shared chat sample links are available: - Normal/Expected behavior: https://chatgpt.com/share/6a604011-8db0-83e8-b3bb-9399a4460919 https://chatgpt.com/share/6a604011-8db0-83e8-b3bb-9399a4460919 - 5.6 Pro → 5-5-mini: https://chatgpt.com/share/6a616b12-d9b8-83ee-b1e0-acb29487fe45 https://chatgpt.com/share/6a616b12-d9b8-83ee-b1e0-acb29487fe45 - 5.5 Pro → i-cot: https://chatgpt.com/share/6a603bf3-0590-83ee-9f56-6d8b573d43da https://chatgpt.com/share/6a603bf3-0590-83ee-9f56-6d8b573d43da At the time of writing, there are two objective ways to check: - For chats you own, inspect the backend data in JSON via DevTools. If the chat was shared by someone else, fork it first. You will own that fork. - Inspect the server-side rendered/injected data of a shared chat. ChatGPT injects the shared-chat data frontend state into the HTML. This method is more complex because the data is not provided as normal JSON, but it's a reliable way to retrieve it when the first method is unavailable. In both cases, check the prop value near resolved model slug . Expects i-cot or gpt-5-5-mini or similar. The first method is strongly recommended. The only advantage of the second method is that it does not require login. While it was possible before, requesting a data dump export and inspecting resolved model slug is unavailable now. The property seems to have been removed from exports recently. A comparison is available in this GitHub Gist link https://gist.github.com/ilsubyeega/c76617cd97a071a4ef4dab81f2d0cc76 . 1 r1 ChatGPT Plan | Pro. https://chatgpt.com/plans/pro/ visited 2026-07-30 archived https://archive.md/H0nR2 reports rreports Disclaimer: Some portion of texts are not exact; some may truncated, machine-translated etc. @paulscu1 06-14 : Just tried OpenAI’s web UI, I’m shocked at just how bad the model routing is. I’m paying $100 a month, and I can’t get more than one reliable response from 5.5 pro before I start getting routed to like 5.3 mini. It’s absurd product design, I will pay more, just please please fix this it doesn’t work X Post https://x.com/paulscu1/status/2065886816238985490 @Mortid X 06-18 : ChatGpt Pro extended 5.5 GPT-5.3-mini model. This is some kind of bug. On my phone, everything works fine for me. I hope the @OpenAI team fixes this bug. Watch the video for proof: video X Post https://x.com/Mortid X/status/2067343110024085958 @ludicrouslytrue 07-21 : My web chat silently falls back to 5.5 mini? I'm on the 20x Pro plan @thsottiaux I thought codex and web limits weren't shared, oh well. video X Post https://x.com/ludicrouslytrue/status/2079532482928718203 @sizumita 07-23 : Why is it a 5.5-mini when it's a Pro model? X Post https://x.com/sizumita/status/2080269517176144169 @ machi47 07-24 : across all of my GPT chats on the Pro plan, when using GPT 5.6 Sol Pro, it routes to GPT 5.5 mini. What is happening @thsottiaux ???? no notices of degredation on the app itself. shows pro level, but does not reason at all, jus instantly spews garbage? X Post https://x.com/ machi47/status/2080394309715849434 @q8jx 07-24 : does your gpt 5.6 pro say it's gpt 5.5 mini too? X Post https://x.com/q8jx /status/2080417495765184514 @RealWilliamHsu 07-24 : You guys doing great but ChatGPT with 5.6 Pro returned replies using GPT 5.5-mini, how's that a genuine reply to the users?? X Post https://x.com/RealWilliamHsu/status/2080469030499172453 @kumataro x 07-26 : @OpenAI Support rejected my refund request, saying it could not confirm that my Pro requests were handled by GPT-5.5 mini. I now have structured data returned by ChatGPT's own web backend showing exactly why this requires an engineering review. I selected Pro. The conversation data shows: default model slug: gpt-5-6-pro But two separate requests show: resolved model slug: gpt-5-5-mini This is machine-readable metadata—not a guess based on response speed or AI self-identification. I obtained the JSON directly from the response payload in my browser's DevTools Network panel while loading the original ChatGPT conversation. The conversation ID in the page URL exactly matches the JSON. I also preserved the original file and its SHA-256 hash. OpenAI's Help Center says Pro is powered by GPT-5.6 Sol Pro. It documents GPT-5.5 Instant mini as an Instant-limit fallback—not as a fallback for selected Pro requests. @OpenAI, please explain what resolved model slug means and check the actual serving model for my two request IDs. If Pro served them, why did your backend return mini? If mini served them, acknowledge the routing issue and reconsider the refund. X Post https://x.com/kumataro x/status/2081060343955296414 @whatfordinner12 07-28 : I pay $100/month for ChatGPT Pro. The UI showed GPT-5.6 Pro. Pro quota was charged. But the request was routed to GPT-5.5-mini. But after I submitted screenshots, chat links, and quota evidence, Support twice dismissed it as “AI hallucination.” 1/5 X Post https://x.com/whatfordinner12/status/2082056790091743508 @TuringMeme 07-28 : I’m a ChatGPT Pro user. Today I noticed that selecting GPT-5.5 Pro automatically switched my conversations to GPT-5.5 Mini. Support explained that Mini is the fallback after reaching the Pro usage limit. I understand having a fallback model, but the problem is: the quality difference is obvious, and there is no clear indication of when Pro access will be restored. Please show the limit status and reset time. Transparency matters for Pro users. X Post https://x.com/TuringMeme/statu