# Why is Haiku smarter than Opus sometimes? — retained record

> Retained reddit record from reddit:ClaudeCode.

Canonical: https://highsignal.app/data/records/21ed4170ffb6d435

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:ClaudeCode)  published 2026-11-01 18:52 ingested 2026-11-02 02:41

[original source ↗](https://www.reddit.com/r/ClaudeCode/comments/1xlb3wl/why_is_haiku_smarter_than_opus_sometimes/)

## retained text



```
I do a lot of retro game dev, and I've noticed changing what model I choose doesn't seem to affect the quality of work much, and just affects how much tokens it uses to do the task. Confused, I did a couple of side by side tests, and one really stood out.

"Look up a pixel art retro gaming tileset online, generate a map for a video game level, and present it in chat as a png"  
I set all 5.5 models to medium effort, and waited.

Haiku: After 2 minutes, it chose an ugly tileset, but otherwise presented a completely usable sidescrolling map.

Sonnet: After 5 minutes, it chose a different ugly tileset, and presented a low quality top down map I'd never use, as the paths didn't connect, and tiles seemed randomly placed.

Opus: after 15 minutes, it failed completely. It said didn't have the tools it needed, and it wanted to install Claude Chrome for some reason.

Have others noticed how "smarter" models will randomly be a lot less intelligent?

  
Edit: People are misunderstanding my post, so I should elaborate.   
  
This is an easy task: This test is simple on purpose, for the sake of comparison. It's not a reflection of the work I do day to day. If it can't do a simple task without overthinking it, that has huge knock on effects when it's given a massive task list

Opus didn't fail, you just didn't give it the tools it needed: Opus DOES fail, all the time. Opus has never once a single time in the many months I've been using it produced a usable result in one shot. No model has. Game development is a complex web of interconnected parts, and game dev is way outside it's training data, so it has to do a lot of guessing. I always have to go in and manually untangle whatever it didn't understand. If it could one-shot a task autonomously, it'd be worth the usage. But if it's gonna get hung up and make me babysit it, then it's not worth it. 

Opus still better: Genuine question, what are y'all even doing with Opus? Because everything I do with Claude in any capacity requires some level of creativity and a subjective eye, things AI is objectively not good at. Everyone on this subreddit seems to be the business type with a $200 ultra subscription, doing vaguely corporate sounding things. I seem to the only broke artist trying to make their work less tedious. I'm not mad about it, but I am very alienated.

*Processing img ss0zxe9ds1zh1...*

*Processing img 4es3zeams1zh1...*
```



document fetched 2026-11-02 00:35 · canonical https://reddit.com/r/ClaudeCode/comments/1xlb3wl/why_is_haiku_smarter_than_opus_sometimes
