ChatGPT Retired a Model and Added a Silent Fallback: The 15-Minute Check on Every Prompt You Reuse
On August 26 OpenAI pulled o3 out of ChatGPT and started routing paid users to GPT-5.4 mini when they hit usage limits. Any saved prompt, Project, or custom GPT you tuned on the old model now runs on something else. Here is the 15-minute check that tells you if the output still holds.
What happened. OpenAI's release notes list two changes worth an operator's attention. First, the o3 model was retired from ChatGPT on August 26, 2026, after a 90-day sunset. The change is ChatGPT only; the API is unchanged. Second, GPT-5.4 mini is rolling out. Free and Go users get it through the "Thinking" option in the + menu. For Plus, Pro, and other paid plans, GPT-5.4 mini is now the fallback for GPT-5.4 Thinking when you hit a rate limit.
Why it matters. Two things can now change the quality of your output without telling you.
- Retirement. If you built a custom GPT, a Project, or a saved prompt while o3 was selected, that work runs on a different model today. Usually it is better. Sometimes the format drifts, a number gets rounded differently, or the tone shifts. You will not know until you look.
- Fallback. On a heavy day, a request you sent to GPT-5.4 Thinking may be answered by mini. For a quick email that is fine. For a contract summary, a financial rollup, or a client-facing letter, a thinner model at 4 pm on your busiest day is a real risk.
The 15-minute check.
- List what you reuse. Custom GPTs, Projects, and the 5 to 10 prompts you paste every week. Write them down. Most operators find fewer than a dozen.
- Set the model on purpose. Open each one and pick the model explicitly in the selector. Anything that was pinned to o3 has already been moved for you. Decide where it should live.
- Re-run your best example. Take the one input where you know exactly what good output looks like. Run it again on the current model. Compare format, numbers, and tone side by side. Fix the prompt if anything slipped.
- Learn to spot the fallback. Each reply in ChatGPT shows which model answered it. When a response reads thinner than usual, check the label. If it says mini and the task matters, wait a few minutes or re-run it.
- Put retirements on the calendar. Every model retirement gets a sunset notice. The day one lands, repeat steps 2 and 3. Ten minutes per retirement beats discovering the drift in front of a client.
Paste this to compare old and new output.
Role: You are a quality checker comparing two versions of the same AI output.
Context: Version A was produced by the model I used to rely on. Version B was produced by the current model from the exact same prompt and input. [Paste Version A, then Version B.]
Task: List every difference that would matter to a business owner: changed numbers, missing sections, added claims, format changes, and tone shifts. Ignore harmless wording changes.
Format: A short table with three columns: what changed, which version is better, and whether I need to edit my prompt to fix it.
Constraints: Do not invent differences. If the two are equivalent for business use, say so in one line.
Keep customer data and financials out of this comparison unless you are in a private business workspace. Use a redacted example.
Source: OpenAI Help Center, Model Release Notes (help.openai.com/en/articles/9624314-model-release-notes).