Why did GPT-4 get worse at coding lately

If GPT-4 suddenly can't write the same quality code it handled last month, you're not imagining it. Developers across GitHub, Reddit and Discord are reporting the same thing: GPT-4 that once generated clean Python functi

Why did GPT-4 get worse at coding lately

If GPT-4 suddenly can't write the same quality code it handled last month, you're not imagining it. Developers across GitHub, Reddit and Discord are reporting the same thing: GPT-4 that once generated clean Python functions now produces buggy loops, and code that previously passed tests now throws errors.

What's actually happening

OpenAI continuously updates GPT-4 through what they call "model snapshots." When you use ChatGPT Plus or the API, you're not always hitting the exact same model version. The default GPT-4 you access today might be `gpt-4-0125-preview` or the newer `gpt-4-turbo-2024-04-09`, while three months ago you were using `gpt-4-0613`. Each snapshot gets trained with different priorities, safety filters and optimization targets.

Here's the part that frustrates developers most: OpenAI doesn't optimize purely for coding ability. They balance multiple objectives including safety, refusal rates, response length, token efficiency and general conversation quality. When they adjust one dial, coding performance sometimes drops. The May 2024 updates, for instance, made GPT-4 more concise but also more cautious about generating complete code blocks without excessive warnings.

You'll see this manifest as incomplete functions, added comments like "// Add error handling here" instead of actual error handling, or the model suddenly refusing to write database queries it handled easily before. Another common sign: GPT-4 now explaining what code should do rather than just writing it, padding responses with disclaimers that eat into the useful output.

How to fix it

1. Lock to a specific model snapshot

If you're using the API, specify the exact model version in your requests instead of using the rolling `gpt-4` alias. Pin to `gpt-4-0613` or whichever snapshot worked well for your use case. Find the full model list at `platform.openai.com/docs/models`.

In your API call:

```

model: "gpt-4-0613"

```

ChatGPT Plus users can't pick snapshots directly, but you can select GPT-4 vs GPT-4 Turbo in the model dropdown at the top of a new chat.

2. Adjust your system prompt

Add explicit instructions that override the model's new cautious defaults:

"Provide complete, working code without placeholders. Do not add comments explaining what I should implement myself. Include all error handling and edge cases in the code you generate."

This often bypasses the "helpful" behaviour that leaves gaps in the output.

3. Use temperature settings strategically

Lower the temperature to 0.3 or 0.4 through the API settings. The newer snapshots respond better to deterministic prompting for code generation. Higher temperatures (0.7+) now produce more variable quality than they did six months ago.

4. Check if you've hit model degradation periods

OpenAI sometimes runs traffic through different model versions during high load. If coding quality tanks during US evening hours but works fine at 3 AM EST, you're likely hitting capacity management. The ChatGPT not working troubleshooting guide covers how to verify if you're experiencing service degradation.

If that doesn't work

Document specific examples where the same prompt produced better code previously. Include the exact model version, timestamp, and both outputs. Submit this through the feedback thumbs-down button in ChatGPT or file a report at `help.openai.com`.

Response times vary wildly, from 48 hours to never, but detailed regression reports do influence model team priorities. Check the model behavior changed page for the template OpenAI's support team actually reviews.

For API users: monitor the `model` field in your response objects. If you requested `gpt-4-0613` but received a different snapshot, that's a billing issue worth escalating through OpenAI's official support channels.

Questions people actually ask

Q: Did OpenAI make GPT-4 dumber to save money?

A: No evidence supports this. More likely they optimized for different metrics that inadvertently reduced coding performance. Token efficiency improvements often hurt code quality.

Q: Is Claude or GitHub Copilot better now?

A: Many developers report Claude 3.5 Sonnet handles code generation more consistently as of late 2024, though it has different strengths. Copilot uses different models entirely.

Q: Will GPT-4 coding ability improve again?

A: OpenAI releases new snapshots every 2-3 months. Some improve coding, others don't. Pinning to a specific version gives you consistency instead of gambling on updates.

Q: Can I access older, better GPT-4 versions?

A: Through the API, yes. `gpt-4-0613` remains available for now. ChatGPT Plus users can't access deprecated snapshots once OpenAI retires them.

What to remember

  • GPT-4 isn't a single fixed model; it's rolling snapshots with different capabilities
  • Lock API calls to specific model versions like `gpt-4-0613` to maintain consistent quality
  • Add explicit "no placeholders, complete code" instructions to your system prompts
  • Coding quality varies by time of day during high platform load
  • Document regressions with specific examples when reporting to OpenAI

---

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

Related help