I was away from my computer for two weeks, and after coming back and reading the latest discussions on HN about coding assistants (models, harnesses), I felt very out of the loop. My normal process would have been to keep reading and figure out the latest and greatest from people's comments, but I wanted to try and automate this process.
Basically the goal is to get a quick overview over which coding models are popular on HN. A next iteration could also scan for harnesses that people use, or info on self-hosting or hardware setups.
I wrote a short intro on the page about the pipeline that collects and analyzes the data, but feel free to ask for more details or check the Google Sheet for more info.
One thing for sure is that while Claude is currently taking the #1 spot in mentions, it carries a lot of negative sentiment due to API pricing policies and frequent server downtime. On the other hand, the runner-up, GPT-5.5, actually seems to have more positive feedback.
Personally, my experience with Codex wasn't as good as with Claude Code (Codex freezes on Windows more often than you'd expect), so this is a bit surprising. That said, the more defensive GPT is definitely better in terms of sheer code-writing capability. However, GPT actually has quite a few issues with text corruption when generating in Korean or Chinese—something English-speaking users probably don't notice. In terms of model capabilities, when given the same agent.md (CLAUDE.md) file, I think GPT is better at writing code, while Claude is better at writing text during code reviews.
Looking at the bottom right, Qwen and DeepSeek are open-source, so they are largely mentioned in the context of guarding against vendor lock-in, which drives positive sentiment. Considering that Hacker News occasionally shows negative sentiment toward China, the fact that they are viewed this positively—unlike US models—shows that being open-source is a massive advantage in itself.
Anyway, one thing for sure is that Gemini is pretty much unusable.
Gemini is not at all unusable. It is quite usable for the tasks it excels at - to the point that it is the top pick for many tasks and I spend more money there than elsewhere. On the other hand it responds quite differently from the other major models - so that claude and gpt on one hand are similar and gemini requires a different approach. In my opinion people who think gemini is worthless have not learned how to prompt it correctly. Again, it's intuitive and watching concrete response difference due to small input changes, but if I had to summarize it shows its google books / google scholar roots.
I have started experimenting with qwen more than deepseek, but I have not had good results yet. Given the good press I presume I will learn how to interact with it for better results.
Curious if others have similar experiences in comparing models usefully, or if most don't bother with this, or do something else? I mainly use models for highly focused specialty tasks, so this fine tuning makes the difference between usable and unusable. I don't yet have the luxury of defining my preferred workflow and finding the tool for the task. Everything just breaks almost immediately if I try to shoehorn into my preferred flow.
And what use cases do you think it’s best suited for?
Its really a cost effective model.
They are cheaper! All signals point to them staying cheaper because they are built more sustainably. Also, some of the latest entries can run on 1 GPU! Literally available at your desktop where there can be no service interruptions. Not even network latency. People are one and few shotting little games for 0 dollars because they bought a GPU to play video games this year. To me that's an unbeatable value. Once the tooling catches up and a few more model releases, it could change everything completely.
Of course, when I tried it on something else it rewrote every line in the file for no good reason, applied changes directly when I told it just to plan, etc.
So maybe it has one strength.
If I have a task that requires parsing through swathes of irregular data that traditional ml would choke on (or require an intermediate training step ala bigquery), I have gotten much better results from Gemini than the other two.
Ha! I find that Gemini is quite useful - if only because I am forced to use it (on my personal projects) because it's the only one that has unlimited interaction for "free"
It has its limitations, yes, but so does Claude (which I am leaning on too heavily at work at the moment)
Subjectively, it seemed like DeepSeek V4 Pro had the highest hype/performance ratio (meaning high hype for lower performance). Whereas MiMo V2.5 Pro didn't get much attention despite being the top dog in the open weights world, not even an honorable mention in your chart :( ...
Searching for it on HN shows very few results, that's why it's not showing up in the analysis yet. But it might in the future, once it gains traction.
I'll keep an eye on it, thanks for bringing it up!
I am upset because now anthropic, openai, meta, etc will continue their smear campaigns here. But I am also happy because it will make HN less useful when they do.
Everything is a give and take I guess. Excited to see where the equilibrium sits
What I want is more fully open models where everything is shared. Data, training algorithms, weights. That way we can figure out if we should trust it.
I think it's also unfair to say their success is solely due to stealing data. They are contributing a lot of advances to the literature about what they are doing. The proof is in the results we have 27b models you can vibe code with. Not 1t+
It's murky sure. But there are smear campaigns about how people can't trust China too. There's some truth to that too but we can't trust the US either so local models are an interesting way for China to offer us some level of sovereignty.
It's actually pretty difficult to find a good comparison model because there isn't one. Again, a 14/28 cent in/out model, ignoring cache, it scores just below GPT 5.4 Mini-xhigh (75/450) and Gemini 3 Flash (50/300) in intelligence. It's similar to Gemma 4 31B in some metrics (13/38) including cost, so it's not completely unheard of, but it's pretty notable that virtually everything else in the same region in most benchmarks are going to cost at least 5 times more (much, much more in very output-heavy contexts)
I wouldn't use Gemini 3 Flash or GPT 5.4 mini for anything except the most trivial work, although both are useful for basic exploratory work.
So I'm using a heavy model for the bulk of the work and the cost of that so far outweighs the light model that the light model cost is effectively irrelevant.
If one likes a model then it's capable of one-shotting entire apps.
Otherwise it's "only suitable for the most trivial tasks".
Never in between.
Personally my opinion in this regard is highly consistent over time.
Edit: Done
It's way too important a piece of information not to have it visible.
And it's probably a good idea to create a list of model release dates, so older comments can't accidentally map to models that weren't released yet.
The context would be really nice to have, but reading the comments myself, it often just isn't very clear what exactly users are building or which programming language they are using.
I think analyzing more comments is promising. If you get enough data, you can generalize across use cases and get more meaningful ratings. The obvious lever is including more posts, although it might hit diminishing returns. I'll play around with it.
For the context, I want to try giving Gemini a "scratch pad", where it can note down strengths and weaknesses per model that it finds in the comments. Something like "some users say that model x is good for writing tests". Then on each run, I let it update the scratch pad and publish the results as more of a qualitative analysis.
For the wording, I'd like to keep a certain amount of click bait, sorry ;)
I saw you're using Gemini for the sentiment rating (which I guess you picked because it's not often mentioned and thus "neutral"? lol)
But would be interesting to get more details overall
I've been experimenting with the 26B-A4B model with some surprisingly good results (both in inference speed and code quality — 15 tok/s, flying along!), vs my last few experiments with Devstral 24B. Not sure whether I can fit that 35B Qwen model everybody's so keen on, on my 32GB unified RAM.
However I think I may be in the minority of HN commenters exploring models for local inference.
kimi...?
The technical abilities and usage are derived from the commenters usage reflections.