Hi everyone. I'm hoping to get some community insight on a specific performance issue I've been observing today. As the title says, I've been getting consistently worse code output from Claude Sonnet compared to Claude Haiku for a similar set of programming tasks.
To give some context, my workflow involves generating boilerplate API client code and small utility functions. Normally, Sonnet is my go-to for this—it provides more robust and well-documented snippets. Today, however, its suggestions have been oddly off. For example, when asked for a simple function to parse and validate a date string, Sonnet's code had a logic error in the conditional check and used a deprecated method. Haiku's version, while more basic, was functionally correct and ran without issue.
I'm trying to rule out my own prompting as the variable. I've used the same prompt structure, specifying language and constraints, across both models. The tasks aren't exceptionally complex, so I wouldn't expect Haiku to outperform Sonnet.
Has anyone else noticed a dip in Sonnet's coding accuracy or clarity recently? I'm wondering if there was a model update or if there are known temporary performance fluctuations between the different Claude models. I'm also curious if certain types of prompts are more susceptible to this. I value vendor neutrality, so I'm just looking for practical, shared experiences here.
Keep it civil, keep it real
I've seen some weird regressions on specific tasks before, but not across the board. What language was the date function for?
Can you share the exact prompts you used? Even minor wording can sometimes trigger a different reasoning path, especially on off days. I'd also check your temperature settings. If you're using a high value, try bringing it down for code generation.
Benchmarks or bust.
That's a good point about ruling out your own prompting as the variable, but I'm curious about the setup. Have you checked if you're hitting a specific geographic server cluster or instance that might have been updated earlier? I've seen temporary "model drift" on individual instances before where one seems off for a few hours.
It's rare for a consistent dip across the board. Could it be the complexity of the tasks? Sometimes the larger models can overcomplicate a simple request where Haiku's more direct approach works perfectly.
What language were you working in? I'd be interested to hear if others are seeing similar things in the same stack.