GPT-4 responses getting worse lately why

You're not imagining it — GPT-4 really does feel different. Your prompts that worked perfectly two months ago now produce generic responses, incomplete code or weird refusals. This happens because OpenAI continuously upd

GPT-4 responses getting worse lately why

You're not imagining it — GPT-4 really does feel different. Your prompts that worked perfectly two months ago now produce generic responses, incomplete code or weird refusals. This happens because OpenAI continuously updates the models behind the scenes, and what you're calling "GPT-4" today might be a different version than what you used last month.

What's actually happening

OpenAI doesn't keep GPT-4 frozen in time. They release new snapshots with updated training data, safety filters and performance tweaks. When you select "GPT-4" in ChatGPT, you're typically getting the latest default version — currently `gpt-4-turbo-2024-04-09` for most users, though this changes without announcement.

These updates cause three common problems. First, response style shifts. A model trained with newer RLHF (reinforcement learning from human feedback) data might produce shorter answers, avoid certain topics more aggressively, or structure responses differently. Second, context handling changes. Newer versions sometimes compress earlier conversation turns differently, making the model "forget" details you referenced earlier in long chats. Third, capability regression. OpenAI occasionally prioritises safety or speed over raw performance, meaning tasks like complex code generation or nuanced analysis may produce worse results than older snapshots.

You'll also hit this if you're using the API. The `gpt-4` endpoint automatically points to whatever OpenAI considers the current stable release. If you built your application against `gpt-4-0613` behaviour and suddenly everything breaks, it's because OpenAI moved the pointer to a newer snapshot. Check your API logs for the actual model identifier in the response metadata — that tells you which version processed your request.

How to fix it

1. Pin to a specific snapshot

If you're using the API, replace `gpt-4` with an exact model ID like `gpt-4-0613` or `gpt-4-1106-preview`. Find the full list at platform.openai.com/docs/models. This locks your requests to one version until OpenAI deprecates it (usually 6-12 months notice). In ChatGPT Plus, you can't pin versions directly, but you can check what changed in model behaviour to understand current defaults.

2. Rewrite your prompts

Newer models respond differently to the same instructions. If "write detailed code comments" now produces brief notes, try "add multi-line docstrings explaining the purpose, parameters and return value for every function". If creative writing feels flat, explicitly request "use varied sentence structure and vivid sensory details" instead of assuming the model remembers your style from earlier in the conversation.

3. Test with different models

In ChatGPT, switch between GPT-4 and GPT-4o using the model selector above the input box. GPT-4o is faster and handles multimodal tasks better, but sometimes produces less detailed text responses. For API users, test `gpt-4-turbo-preview` against older snapshots to find which version handles your use case best. Don't assume the newest model is always better for your specific task.

4. Reduce conversation length

If responses deteriorate after 10-15 exchanges, start a new chat. Newer models compress context aggressively to stay within token limits, which can distort earlier instructions. For API use, implement conversation summarization every 8-10 turns instead of sending the full history.

If that doesn't work

Document exactly what changed. Save example outputs from "before" (if you have chat history) and "after". Note the date when behaviour shifted — this helps identify which model update caused it. Then contact OpenAI support through help.openai.com with:

  • Your account email and organisation ID (for API users)
  • Specific example prompts and unwanted outputs
  • The model ID from API response logs, if applicable
  • The date you first noticed the change

API users get faster support through the platform dashboard. ChatGPT Plus users should set expectations for 3-7 day response times. OpenAI rarely rolls back model changes based on individual reports, but they do track patterns. If you're experiencing actual ChatGPT system failures rather than quality changes, that's a separate issue requiring different troubleshooting.

Questions people actually ask

Q: Can I access older GPT-4 versions in ChatGPT?

A: No. ChatGPT Plus only offers the current default versions. API users can specify older snapshots until OpenAI deprecates them.

Q: Does model quality drop when servers are busy?

A: No. OpenAI doesn't reduce model capability during high load. You might hit rate limits or slower response times, but the actual model version stays the same.

Q: Will complaining make OpenAI revert changes?

A: Individual complaints rarely trigger rollbacks. However, widespread reports of specific regressions (like the December 2023 "lazy GPT-4" issue) sometimes prompt OpenAI to release updated versions addressing those concerns.

What to remember

  • "GPT-4" is a moving target — the underlying version changes regularly without notification
  • Pin to specific snapshots via API (`gpt-4-0613`) if you need consistent behaviour
  • Rewrite prompts for newer models instead of assuming old instructions still work
  • Start new conversations after 10-15 exchanges to avoid context compression issues
  • Document changes with examples and dates when reporting problems to OpenAI

---

*openai-support.com is an independent resource, not affiliated with OpenAI Inc.*

Related help