Why Does ChatGPT Seem Worse Now? Check What Changed

When ChatGPT answers start feeling less useful or less sharp, there are three things to check before concluding the product has gotten worse: which model your session is using, how much context the session has, and whether the task you are asking about has changed. Each of these has a practical fix, and one of them usually explains the problem.

Tools like fenbs keep a consistent task record outside the conversation, so sessions missing context can be recovered without retyping, and comparing answers across assistants is straightforward when you have the same starting notes ready.

Step 1: Check which model your session is using

ChatGPT does not always use the same model in every conversation. On the free tier, you may be switched to a less capable model when the more capable one is busy or you have hit your daily limit. On paid plans, the model selector near the compose area shows which model is active. If it shows a different model than you expect, that is the most likely cause of weaker answers.

OpenAI also updates models regularly. A model update sometimes changes the response style, makes the model more cautious, or adjusts how it handles certain task types. If your account shows the same model name but the answers feel different, the model itself may have been updated. Check the OpenAI changelog↗ (opens in a new tab) for recent model notes.

Step 2: Check what context the session has

A new conversation starts with no memory of previous sessions. If you are asking ChatGPT to continue work from an earlier chat without re-establishing the context — the task notes, the rules, the decisions already made — the answers will reflect that missing information. A vague or general answer is often the model doing its best with an underspecified prompt, not the model getting worse.

Check whether you are starting each session with the relevant background. If you have a project that spans multiple conversations, the session that gets the full context typically produces better answers than one that does not. Memory features (when active) help, but they are selective and do not replace an explicit context at the start of a session.

Step 3: Compare on the same task

Before concluding ChatGPT has gotten worse overall, test it on the same prompt from a previous session where it performed well. Use the exact same wording. If the newer output is worse, that is a meaningful data point. If it is similar, the problem may be that you are now asking for something harder or different from what you expected.

Also run the same prompt in a second assistant to compare. If both produce similarly weak results, the task is harder than it looks or the prompt needs work. If one handles it clearly better, you have identified a genuine capability difference worth acting on.

Common reasons ChatGPT answers feel different

• You are on the free tier and using a less capable model than before. Check the model selector.
• A model update changed how the model handles requests in your task type — more caution, different tone, shorter answers.
• The session is missing context that earlier sessions had. The model is not worse; it knows less about your work.
• Your task has gotten harder or more specific, and the quality gap relative to the task is more visible.
• ChatGPT Work features changed: if you use a team account, plan updates can change what is available.

When switching assistants makes sense

After testing context and model, if a specific task type consistently produces better results in a different assistant, switching for that task is a reasonable choice. Claude, Gemini and Copilot have different strengths, and a consistent difference on writing tasks, coding tasks or research is a signal worth acting on.

Switching assistants is low-friction when the project context lives outside the model. The next assistant reads the same tasks and rules from fenbs and starts from the same place. What you do not have to do is re-explain the project from scratch.

For a direct comparison of ChatGPT and Claude on writing tasks, see Claude vs ChatGPT. For alternatives broadly, see ChatGPT alternatives.

One practical note: if ChatGPT consistently underperforms on a task type, record that finding somewhere you can revisit it. Model updates are frequent, and a task that one version handles poorly may be handled well in the next. Noting the comparison and date means you can re-test when a new version lands rather than assuming the gap is permanent.

How to do this with fenbs

1. Create a free account and write your current project context in fenbs.
2. Record the task details and any decisions or rules the assistant should follow.
3. Connect ChatGPT from Integrations and ask it to read the project context before each session.
4. When you test a different assistant on the same task, use the same fenbs context so the comparison is fair.
5. Record which assistant worked better for which task type as a context note so the next session starts from that finding.