GPT-4 responses getting worse and shorter — why OpenAI nerfed the model and what to do
You've noticed GPT-4 suddenly giving you lazy three-sentence answers when it used to write detailed paragraphs. Your prompts that worked perfectly two months ago now get half-finished responses or generic brush-offs. You
GPT-4 responses getting worse and shorter — why OpenAI nerfed the model and what to do
You've noticed GPT-4 suddenly giving you lazy three-sentence answers when it used to write detailed paragraphs. Your prompts that worked perfectly two months ago now get half-finished responses or generic brush-offs. You're not imagining it — GPT-4's behavior has shifted, and thousands of users are reporting the same drop in output quality.
What's actually happening
OpenAI didn't intentionally "nerf" GPT-4, but the model's behavior has changed due to several overlapping factors. First, the default model you're using likely switched. If you're on ChatGPT Plus, the standard "GPT-4" option now points to different model versions depending on load — sometimes GPT-4-turbo, sometimes GPT-4o, sometimes optimized variants designed for speed over thoroughness.
Second, OpenAI implemented stricter system instructions around verbosity starting in mid-2024. Internal guidelines now push models toward conciseness by default. What you're experiencing as "worse" is the model following instructions to be brief unless you explicitly ask for detail. The training data cutoffs also shifted — newer model iterations prioritize recent conversation patterns where users often request shorter answers.
Third, rate limiting affects response depth. When you're approaching your usage caps (25 messages per 3 hours on Plus, 80 messages per 3 hours on ChatGPT Pro), the system may route you to faster, less computationally expensive responses. You won't get a warning — the answers just get thinner.
How to fix it
1. Force detailed responses in every prompt
Add this exact instruction at the end of your prompt: "Provide a comprehensive answer with specific examples and detailed explanations. Do not summarize."
For coding tasks, specify: "Write the complete code with all functions implemented. Do not use placeholder comments."
2. Switch to GPT-4o explicitly
On ChatGPT web interface, click the model selector at the top of the chat window. Choose "GPT-4o" instead of the default "ChatGPT" or "GPT-4". GPT-4o maintains longer context and more consistent output quality than the load-balanced default option.
On API calls, specify `"model": "gpt-4o"` in your request instead of `"model": "gpt-4"`.
3. Use the regenerate button strategically
When you get a short response, click "Regenerate response" immediately. The second attempt often pulls from a different model instance with different verbosity settings. If that fails, reply with: "This response is too brief. Please expand each point with specific details and examples."
4. Check your usage tier
Go to platform.openai.com/settings/organization/limits (for API) or your ChatGPT account settings to see your current rate limits. If you're hitting caps, responses degrade. Upgrade to ChatGPT Pro for higher limits, or for API users, request a tier increase at platform.openai.com/settings/organization/billing.
5. Create a custom instruction set
In ChatGPT, go to Settings (bottom left) → Personalization → Custom instructions. In the second box ("How would you like ChatGPT to respond?"), add: "Always provide thorough, detailed responses unless I specifically ask for a summary. Include all relevant steps, code, or examples without placeholders."
If that doesn't work
Document three specific examples where responses failed your expectations. Include the exact prompt, the model version you selected, the timestamp, and the unsatisfactory response. Email help.openai.com with subject line "Model response quality degradation — GPT-4". Typical response time is 3-5 business days.
For API users experiencing systematic issues, submit a report through platform.openai.com/account/support with your organization ID and specific request IDs from your logs. Include rate limit info from your dashboard.
Questions people actually ask
Q: Did OpenAI downgrade Plus subscribers to push people toward Pro?
A: No evidence supports this. Response variations come from load balancing and model routing, not deliberate throttling by subscription tier. Pro users report the same inconsistencies.
Q: Why does the same prompt give different response lengths each time?
A: ChatGPT routes requests across multiple model instances with different system configurations. Temperature settings, system instructions, and the specific model variant can vary between requests even when you select the same model name.
Q: Will switching to API instead of ChatGPT Plus give more consistent results?
A: Yes, but only if you specify exact model versions like `gpt-4o-2024-11-20` instead of `gpt-4`. API requests with pinned versions eliminate load-balancing variations you see on the web interface.
What to remember
- Add explicit verbosity instructions to every prompt requiring detailed output
- Select "GPT-4o" specifically instead of the default auto-routing option
- Monitor your rate limits at platform.openai.com — hitting caps degrades response quality
- Use custom instructions to set baseline expectations for response thoroughness
- Pin exact model versions in API calls to avoid routing inconsistencies
---
*openai-support.com is an independent resource, not affiliated with OpenAI Inc.*