A recent blog post titled "Why does Opus 5 feel worse to work with?" hit the front page of Hacker News with nearly 1,000 upvotes. That's not a typo — nearly a thousand developers upvoted a post about an AI model feeling worse than its predecessor. The discussion thread was one of the most engaged HN conversations of the week.
This isn't an isolated complaint. Across developer forums, Twitter threads, and Discord channels, there's a growing chorus of users reporting that newer, theoretically better AI models feel worse in practice. What's going on?
The Benchmark vs. Experience Gap
AI model developers optimize for benchmarks. Benchmarks are measurable, comparable, and marketable. When a new model scores higher on coding tests, math problems, and reasoning challenges, that's a clear improvement.
But benchmarks measure capability, not experience.
- Capability is what a model can do in the best case
- Experience is what it feels like to work with in the average case
These are not the same thing. A model can be more capable and simultaneously worse to work with. Here's how.
The Specific Complaints About Opus 5
The blog post and HN discussion identified several patterns:
1. Over-Confidence
Opus 5 tends to present solutions with more confidence, even when it's wrong. Earlier models were more likely to hedge, which paradoxically made them easier to work with — when a model says "I think this might work, but let me verify," you know to check. When a model says "Here's the solution," you're more likely to trust it and move on.
The result: more confident wrong answers that are harder to catch.
2. Verbosity Without Substance
Users reported that Opus 5 generates longer responses that don't necessarily contain more useful information. This is a common failure mode when models are trained to "be helpful" — they learn that longer responses correlate with higher satisfaction ratings, even when the extra length adds no value.
3. Reduced Code Quality on Edge Cases
While benchmark scores went up (likely because benchmarks test common cases), users reported that edge case handling got worse. This makes sense: if you optimize for the cases that appear in benchmarks, you may de-optimize for the long tail of unusual situations that real developers encounter.
4. Personality Shift
This is the hardest to quantify but the most commonly mentioned: Opus 5 just feels different. The tone, the approach to problems, the way it structures responses — something changed in a way that users don't like.
Why This Happens: The Optimization Trap
The fundamental issue is that AI companies are optimizing for metrics that don't fully capture what users want. When you train a model to maximize benchmark scores, you get a model that's good at benchmarks. When you train it to maximize user satisfaction ratings, you get a model that's satisfying in the moment (which may mean being agreeable rather than correct).
Neither of these necessarily produces a model that's good to work with over a long session on a real project.
The Goodhart's Law Problem
When a measure becomes a target, it ceases to be a good measure.
AI benchmarks are the textbook example. Once a benchmark becomes the target that models are optimized against, it stops measuring real-world performance. The benchmark scores go up, but real-world utility doesn't — or goes down.
What Users Actually Want
Reading through the HN discussion, what developers want from an AI coding assistant is remarkably consistent:
- Honesty about uncertainty: Say "I don't know" or "I'm not sure" when appropriate
- Conciseness: Don't pad responses with fluff
- Code that actually works: Not code that looks like it should work
- Consistency: Don't give different answers to the same question on different days
- Awareness of context: Remember what we discussed 10 minutes ago
None of these are well-captured by standard benchmarks. They're qualitative, subjective, and hard to measure at scale. But they're what matters.
The Implications for AI Development
The Measurement Problem
If we can't measure what users actually want, we can't optimize for it. This is the core challenge facing AI model development. The industry needs better ways to evaluate models on real-world experience, not just benchmark performance.
The Versioning Dilemma
When a new model is released that scores higher on benchmarks but feels worse to use, users are put in a difficult position. They can either:
- Use the new model (better benchmarks, worse experience)
- Stick with the old model (worse benchmarks, better experience)
- Switch providers entirely
This is why we're seeing users hold onto older models, use multiple providers, and increasingly express frustration with the upgrade treadmill.
The Market Signal
The fact that nearly 1,000 HN users upvoted a post about a model feeling worse is a powerful market signal. It means this isn't a fringe complaint — it's a widespread sentiment that resonated deeply with the developer community.
Companies that listen to this signal and optimize for experience, not just benchmarks, will have a real competitive advantage.
What You Can Do
If you're a developer working with AI models:
- Report your experience: Feedback channels exist — use them
- Don't upgrade blindly: Test new models on your actual work, not just on benchmarks
- Use multiple models: Different models have different strengths; don't be locked into one
- Track your own metrics: Keep notes on which models work best for your specific tasks
If you're building AI products:
- Measure experience, not just capability: User satisfaction ≠ benchmark scores
- Respect the long tail: Edge cases matter more than benchmarks suggest
- Don't ship a model that feels worse: No amount of benchmark improvement justifies degrading user experience
Conclusion
The question "why does Opus 5 feel worse?" is actually the question "why do our metrics for AI models fail to capture what users actually care about?" It's a question the entire AI industry needs to answer — because users are noticing the gap, and they're voting with their upvotes.
The original blog post is worth reading, and the HN discussion is one of the best threads on AI model quality I've seen this year.
Top comments (0)