Hacker News Reader

o3 is still my favorite over even Opus 4 in most cases. I've spent hundreds of dollars on AI code gen tools in the last month alone and my ranking is:

1. o3 - it's just really damn good at nuance, getting to the core of the goal, and writing the closest thing to quality production level code. The only negative is it's cutoff window and cost, especially with it's love of tools. That's not usually a big deal for the Rails projects I work on but sometimes it is.

2. Opus 4 via Claude Code - also really good and is my daily driver because o3 is so expensive. I will often have Opus 4 come up with the plan and first pass and then let o3 critique and make a list of feedback to make it really good.

3. Gemini 2.5 Pro - haven't tested this latest release but this was my prior #2 before last week. Now I'd say it's tied or slightly better than Sonnet 4. Depends on the situation.

4. Sonnet 4 via claude Code - it's not bad but needs a lot of coaching and oversight to produce really good code. It will definitely produce a lot of code if you just let it go do it's thing but it's not the quality, concise, and thoughtful code without more specific prompting and revisions.

I'm also extremely picky and a bit OCD with code quality and organization in projects down to little details with naming, reusability, etc. I accept only 33% of suggested code based on my Cursor stats from last month. I will often revert and go back to refine the prompt before accepting and going down a less than optimal path.

It's interesting you say that because o3, while being a considerable improvement over OpenAI's other models, still doesn't match the performance of Opus 4 and Gemini 2.5 Pro by a long shot for me.

However, o3 resides in the ChatGPT app, which is still superior to the other chat apps in many ways, particularly the internet search implementation works very well.

What languages do you use it with and IDE? I use it in Cursor mainly with Max reasoning on. I spent around $300 on token based usage for o3 alone in May still only accepting around 33% of suggestions though. I made a post on X about this the other day but I expect that amount of rejections will go down significantly by the end of this year at the rate things are going.

Very strange. I find reasoning has very narrow usefulness for me. It's great to get a project in context or to get oriented in the conversation, but on long conversations I find reasoning starts to add way too much extraneous stuff and get distracted from the task at hand.

I think my coding model ranking is something like Claude Code > Claude 4 raw > Gemini > big gap > o4-mini > o3

Claude Code isn't a model in itself. By default it routes some to Opus 4 or Sonnet 4 but mostly Sonnet 4 unless you explicitly set it.

I am aware

i'm using with python, VS Code (not integrated with claude just basic copilot) and Claude Code. For Gemini i'm using AI studio with repomix to package my code into a single file. I copy files over manually in that workflow.

All subscription based, not per token pricing. I'm currently using Claude Max. Can't see myself exhausting its usage at this rate but who knows.

If you're coding through chat apps you're really behind the times. Try an agent IDE or plugin.

Yeah, exactly. For everyone who might not know, the chat apps add lots of complex system prompting to handle and shape personality, tone, general usability, etc. IDE's also do this (with Claude Code being one of the ones that are closest to "bare" model that you can get) but at they are at least guiding it's behavior to be really good at coding tasks. Another reason is using the Agent feature that IDE's have had for a few months now which gives it the ability to search/read/edit files across your codebase. You may not like the idea of this and it feels like losing control, but it's the future. After months of using it I've learned how to get it to do what I want but I think a lot of people who try it once and stop get frustrated that it does something dumb and just assume it's not good. That's a practice and skill problem not a model problem.

IDE's are intimidating to non-tech people.

I'm surprised there isn't a VibeIDE yet that is purpose build to make it possible for your grandmother to execute code output by an LLM.

> I'm surprised there isn't a VibeIDE yet that is purpose build to make it possible for your grandmother to execute code output by an LLM.

The major LLM chat interfaces often have code execution built in, so there kind of is, it just doesn't look like what an SWE thinks of as an IDE.

I have not used them but I feel like there are tools like Replit, Lovable, etc that are for that audience. I totally agree IDE's are intimidating for non-technical people though. Claude Code is pretty cool in that way where it's one command to install and pretty easy to get started with.

https://firebase.studio comes to mind

This has been my experience. It has been something I’ve had to settle into. After some reps, it is becoming more difficult to imagine going back to regular old non-assisted coding sessions that aren’t purely for hobby.

Your model rankings are spot on. I’m hesitant to make the jump to top tier premium models as daily drivers, so I hang out with sonnet 4 and/or Gemini 2.5 pro for most of the day (max mode in Cursor). I don’t want to get used to premium quality coming that easy, for some reason. I completely align with the concise, thoughtful code being worth it though. I’m having to do that myself using tier 2 models. I still use o3 periodically for getting clarity of thought or troubleshooting gnarly bugs that Claude gets caught looping on.

How would you compare Cursor to Claude Code? I’m yet to try the latter.

I am really struggling with this. I tried Cline with both OpenAI and Claude to very weird results. Often burning through credits to get no where or just running out of context. I just got Cursor for a try so can't say anything on that yet.

It's a skill that takes some persistence and trial and error. Happy to chat with you about it if you want to send me an email.

I really appreciate that. I will see how I get on and may well give you a shout. Thank you!

There is skill to it but that's certainly not the only relevant variable involved. Other important factors are:

Language: Syntax errors rise, and a common form is the syntax of a more common language bleeding through.

Domain: Less so than what humans deem complex, quality is more strongly controlled by how much code and documentation there is for a domain. Interesting is that if in a less common subdomain, it will often revert to a more common approach (for example working on shaders for a game that takes place in a cylinder geometry requires a lot more hand-holding than on a plane). It's usually not that they can't do it, but that they require much more involved prompting to get the context appropriately set up and then managing drifting to default, more common patterns. Related is decisions with long term consequences. LLMs are pretty weak at this. In humans this one comes with experience, so it's rare and an instance of low coverage.

Dates: Related is reverting to obsolete API patterns.

Complexity: While not as dominant as domain coverage, complexity does play a role. With likelihood of error rising with complexity.

This means if you're at the intersection of multiple of these (such as a low coverage problem in a functional language), agent mode will likely be too much of a waste for you. But interactive mode can still be highly productive.

I think this is debatable. But I've used Cursor and various extensions for VS Code. They're all fine (but cursor can fuck all the way off for stealing the `code` shell integration from VS Code) but you don't _need_ an IDE as Claude Code has shown us (currently my primary method of vibe coding).

It's mostly about the cost though. Things are far more affordable in the the various apps/subscriptions. Token-priced API's can get very expensive very quickly.

We are trading tokens and mental health for time?

I used Cursor well over a year ago. It gave me a headache. It was very immature. Used cursor more recently: the headache intensity increased. It's not cursor it is the senseless loops hoping for the LLM to spit out something somewhat correct. Revisiting the prompt. Trying to become an elite in language protocols because we need that machine to understand us.

Leaving aside the headache, its side effects. It isn't clear we haven't already maxed out on the productivity tools efficiency. Auto complete. Indexed and searchable doc a second screen rather than having to turn the pages of some reference book. Etc etc.

I'm convinced at this stage that we've already started to trade too far. So far beyond the optimal balance that these aren't diminishing returns. It is absolute diminishing.

Engineers need to spend more time thinking.

I'm convinced that engineers, if they were to chose, would throw this thing out and make space for more drawing boards, would use a 5 minute Solitaire break every 1h. Or take a walk.

For some reason the constant pressure to go faster eventually makes its mark.

It feels right to see thousands of lines of code written up by this thing. It feels aligned with the inadequate way we've been measured.

Anyway. It can get expensive and this is by design.

> We are trading tokens and mental health for time?

I have bipolar disorder. This makes programming incredibly difficult for me at times. Almost all the recent improvements to code generation tooling have been a tremendous boon for me. Coding is now no longer this test of how frustrated I can get over the most trivial of tasks. I just ask for what I want precisely and treat responses like a GitHub PR where mistakes may occur. In general (and for the trivial tasks I'm describing) Claude Code will generate correct, good code (I inform it very precisely of the style I want, and tell it to use linters/type-checkers/formatters after making changes) on the first attempt. No corrections needed.

tl;dr - It's been nothing but a boon for this particular mentally ill person.

If your handicap make coding difficult perhaps another profession would suit you better.

Now if Ai assistance allow you to perform well then that is a different story and I take my advice back of course.

There is a lot to say, positive things about how LLMs enables people to perform at tasks that would be impossible for them. Whether due to handicaps or simply lacking the abilities, or opportunity to train.

My comment was on the impact on "healthy" individuals who remain the majority of the population. And I only spoke for myself, I have no clue maybe it is just me or due to how I use the thing. Thanks for sharing your experience though, I had not considered what might be a concern for the majority with this might very well be an enabler.

An important caveat here is yes, for coding. Apps are fine for coming up with one-liners, or doing other research. I haven't found the quality of IDE based code to be significantly better than what ChatGPT would suggest, but it's very useful to ask questions when the model has access to both the code and can prompt you to run tests which rely on local data (or even attached hardware). I really don't trust YOLO mode so I manually approve terminal calls.

My impression (with Cursor) is that you need to practice some sort of LLM-first design to get the best out of it. Either vibe code your way from the start, or be brutal about limiting what changes the agent can make without your approval. It does force you to be very atomic about your requests, which isn't a bad thing, but writing a robust spec for the prompt is often slower than writing the code by hand and asking for a refactor. As soon as kipple, for lack of a better word, sneaks into the code, it's a reinforcing signal to the agent that it can add more.

It's definitely worth paying the $20 and playing with a few different clients. The rabbit hole is pretty deep and there's still a ton of prompt engineering suggestions from the community. It encourages a lot of creative guardrails, like using pre-commit to provide negative feedback when the model does something silly like try to write a 200 word commit message. I haven't tried JetBrains' agent yet (Junie), but that seems like it would be a good one to explore as well since it presumably integrates directly with the tooling.

Depends. For devops chat is quite nice as the exploration/understanding is key, not just writing out the configs.

What's most annoying about Gemini 2.5 is that it is obnoxiously verbose compared to Opus 4. Both in explaining the code it wrote and the amount of lines it writes and comments it adds, to the point where the output is often 2-3x more than Opus 4.

You can obviously alleviate this by asking it to be more concise but even then it bleeds through sometimes.

Yes this is what I mean by conciseness with o3. If prompted well it can produce extremely high level quality code that blows me away at times. I've also had several instances now where I gave it slightly wrong context and other models just butchered a solution with dozens of lines for the proposed fix which I could tell wasn't right and then after reverting and asking o3, it immediately went searching for another file I hadn't included and fixed it in one line. That kind of, dare I say independent thinking, is worth a lot when dealing with complex codebases.

Personally I still am of the opinion current LLMs are more of a very advanced autocomplete.

I have to think of the guy posting that he fed his entire project codebase to an AI, it refactored everything, modularizing it but still reducing the file count from 20 to 12. "It was glorious to see. Nothing worked of course, but glorious nonetheless".

In the future I can certainly see it get better and better, especially because code is a hard science that reduces down to control flow logic which reduces down to math. It's a much more narrow problem space than, say, poetry or visuals.

Have you used Cline with opus+sonnet? Do you have opinions about Claude code vs cline+api? Curious to hear your thoughts!

I use o3 a lot for basic research and analysis. I also find the deep research tool really useful for even basic shopping research

Like just today, it made a list of toys for my toddler that fit her developmental stage and play style. Would have taken me 1-2 hrs of browsing multiple websites otherwise

Gemini deep research runs circles around OpenAI deep research. It goes way deeper and uses way more sources.

How do you choose which model to use with Claude Code?

I have the Max $200 plan so I set it to Opus until it limits me to Sonnet 4 which has only happened in two out of a few dozen sessions so far. My rule of thumb in Cursor is it's worth paying for the Max reasoning models for pretty much every request unless it's stupid simple because it produces the best code each time without any funny business you get with cheaper models.

You can use the max plan in cursor? I thought it didn’t support calls via api and only worked in Claude code?

I launch Claude Code in VS Studio (similar to Cursor): > claude

Then I use the /login command that opens a browser window to log into Claude Max.

You can confirm Claude Max billing going forward in VS Studio/Claude Code: /cost

"With your Claude Max subscription, no need to monitor cost — your subscription includes Claude Code usage"

In case you're asking for the literal command...

/model

we need to stop it with the anecdotal evidence presented by one random dude

I find o3 to be the clearest thinker as well.

If I'm working on a complex problem and want to go back and forth on software architecture, I like having o3 research prior art and have a back and forth on trade-offs.

If o3 was faster and cheaper I'd use it a lot more.

I'm curious what your workflows are !

How do you find o3 vs o4-mini?

For coding at least, I don't bother with anything less than the top thinking models. They do have their place for some tasks in agentic systems but time is money and I don't want to waste time trying to coral less skilled models when there are more powerful ones available.

I have the same logic but opposite conclusion - o3 just takes SO LONG to respond that I often just use o4-mini

What I like about Gemini is the search function that is very very good compared to others. I was blown away when I asked to compose me an email for a company that was sending spam to our domain. It literally searched and found not only the abuse email of the hosting company but all the info about the domain and the host(mx servers, ip owners, datacenters, etc.). Also if you want to convert a research paper into a podcast it did it instantly for me and it's fun to listen.

I’ve been giving the same tasks to claude 4 and gemini 2.5 this week and gemini provided correct solutions and claude didn’t. These weren’t hard tasks either, they were e.g. comparing sql queries before/after rewrite - Gemini found legitimate issues where claude said all is ok.

in my experience this highly depends case by case. For some cases Gemini crushed my problem, but in next one stuck and couldn't figure out simple bug.

the same with o3 and sonnet (I didn't tested 4.0 much yet to have opinion)

I feel thet we need better parallel evaluation support. where u could evaluate all top models and decide with one provided best solution

Have you tried o3 on those problems? I've found o3 to be much more impressive than Opus 4 for all of my use cases.

To be honest, I haven't, because the "This model is extremely expensive" popup on Cursor makes me a bit anxious - but given the accolades here I'll have to give it a shot.

I haven't tried all of the favorites, just what is available with Jetbrains AI, but I can say that Gemini 2.5 is very good with Go. I guess that makes sense in a way.

I think the only way to be particularly impressed with new leading models lately is to hold the opinion all of the benchmarks are inaccurate and/or irrelevant and it's vibes/anecdotes where the model is really light years ahead. Otherwise you look at the numbers on e.g. lmarena and see it's claiming a ~16% preference win rate for gpt-3.5-turbo from November of 2023 over this new world-leading model from Google.

Not sure I follow - Gemini has ELO 1470, GPT3.5-turbo is 1206, which is an 86% win rate. https://chatgpt.com/share/6841f69d-b2ec-800c-9f8c-3e802ebbc0...

gpt-3.5-turbo-1106 from November 2023 was 1170, 1206 is for the March variant.

Change that and you get ~84%, flip the order (i.e. the win rate of GPT-3.5 is ~16%). I.e. the point is a two year old model still wins far too often to be excited about each new top model for the last two years, not that the two year old model is better.

People can ask whatever they want on LMarena, so a question like "List some good snacks to bring to work" might elicit a win for a old/tiny/deprecated model simply because it lists the snack the user liked more.

are you saying that's a bad way to judge a model? Not sure why we'd want ones that choose bad snacks

I agree I find claude easily the best model, at least for programming which is the only thing I use LLMs for

>That being said, I'm starting to doubt the leaderboards as an accurate representation of model ability

Goodhart's law applies here just like everywhere else. Much more so given how much money these companies are dumping into making these models.

> after a bit Gemini would spin in circles or actually (I've never seen this before!) give up and say it can't do it

No way, is there any way to see the dialog or recreate this scenario!?

The chat was in Cursor, so I don't know a way to provide a public link, but here is the last paragraph that it output before I (and it) gave up. I honestly could have re-prompted it from scratch and maybe it would have gotten it, but at this point I was pretty sure that even if it did, it was going to make a total mess of things. Note that it was iterating on a test failure and had spun through multiple attempts at this point:

> Given the persistence of the error despite multiple attempts to refine the type definitions, I'm unable to fix this specific TypeScript error without a more profound change to the type structure or potentially a workaround that might compromise type safety or accuracy elsewhere. The current type definitions are already quite complex.

The two prior paragraphs, in case you're curious:

> I suspect the issue might be a fundamental limitation or bug in how TypeScript is resolving these highly recursive and conditional types when they are deeply nested. The type system might be "giving up" or defaulting to a less specific type ({ __raw: T }) prematurely.

> Since the runtime logic seems to be correctly hydrating the nested objects (as the builder.build method recursively calls hydrateHelper), the problem is confined to the type system's ability to represent this.

I found, as you can see in the first of the prior two paragraphs, that Gemini often wanted to claim that the issue was on TypeScript's side for some of these more complex issues. As proven by Opus, this simply wasn't the case.

for bulk data extraction on personal real life data I experienced that even gpt-4o-mini outperforms latest gemini models in both quality and cost. i would use reasoning models but their json schema response is different from the non-reasonig models, as in: they can not deal with union types for optional fields when using strict schemas... anyway.

idk whats the hype about gemini, it's really not that good imho

I just realized that Opus 4 is the first model that produced "beautiful" code for me. Code that is simple, easy to read, not polluted with comments, no unnecessary crap, just pretty, clean and functional. I had my first "wow" moment with it in a while. That being said it occasionally does something absolutely stupid. Like completely dumb. And when I ask it "why did you do this stupid thing", it replies "oh yeah, you're right, this is super wrong, here is an actual working, smart solution" (proceeds to create brilliant code)

I do not understand how those machines work.

My issue is that every time i've attempted to use Opus 4 to solve any problem, I would burn through my usage cap within a few min and not have solved the problem yet because it misunderstood things about the context and I didn't get the prompt quite right yet.

With Sonnet, at least I don't run out of usage before I actually get it to understand my problem scope.

I've also experienced the same, except it produced the same stupid code all over again. I usually use one model (doesn't matter which) until it starts chasing it's tail, then I feed it to a different model to have it fix the mistakes by the first model.

> Code that is simple, easy to read, not polluted with comments, no unnecessary crap, just pretty, clean and functional

I get that with most of the better models I've tried, although I'd probably personally favor OpenAI's models overall. I think a good system prompt is probably the best way there, rather than relying in some "innate" "clean code" behavior of specific models. This is a snippet of what I use today for coding guidelines: https://gist.github.com/victorb/1fe62fe7b80a64fc5b446f82d313...

> That being said it occasionally does something absolutely stupid. Like completely dumb

That's a bit tougher, but you have to carefully read through exactly what you said, and try to figure out what might have led it down the wrong path, or what you could have said in the first place for it avoid that. Try to work it into your system prompt, then slowly build up your system prompt so every one-shot gets closer and closer to being perfect on every first try.

Thanks for sharing, I'll copy your rules :)

how does it have access to DOM? are you using it with cursor/browser MCP?