HN Reader

Remember when GPT-3 came out and everybody collectively freaked the hell out? That's how I've felt watching the reaction to any of the new model releases lately that make any progress.

I'm honestly not complaining about the model releases, though. Despite their shortcomings, they are extremely useful. I've found Gemini 3 to be an extremely useful learning aid, so as long as I don't blindly trust its output, and if you're trying to learn, you really ought not do that anyways. (Despite what people and benchmarks say, I've already caught some random hallucinations, it still feels like you're likely to run into hallucinations on a regular basis. Not a huge problem, but, you know.)

It feels like every model release has its own little hype cycle. Apparently Claude 4.5 is still climbing to its peak of inflated expectations.

This is the new goalpost now that the "this model is so intelligent that it's sentient and dangerous" AGI hype has died down.

Why are we commenting the Claude subreddit?

1) it’s not impartial

2) it’s useless hype commentary

3) it’s literally astroturfing at this point

Almost every single post on the ClaudeAI subreddit is like this. I use Opus 4.5 in my day to day work life and it has quickly become my main axe for agentic stuff but its output is not a world-shattering divergence from Anthropic's previous, also great iterations. The religious zealotry I see with these things is something else.

I used claude code for a while in the summer, took a vacation from LLMs and I'm trying it out again now. I've heard the same thing about Opus 4.5, but my experience with claude code so far is the same as it was this summer... I guess if you're a casual user don't get too excited?

It's a Misanthropic propaganda forum. They even have Claude agree in the summary of the bot comments:

"The overwhelming consensus in this thread is that OP's fear is justified and Opus represents a terrifying leap in capability. The discussion isn't about if disruption is coming, but how severe it will be and who will survive."

My fellow Romans, I come here not to discuss disruption, but to survive!

There are still a few things missing from all models: taste, shame and ambition. Yes they can write code, but they have no idea what needs does that code solve, what a good UX looks like and what not to ship. Not to mention that they all eventually go down rabbit holes of imaginary problems that cannot be solve (because they’re not real), and do where they will spend eternity unless w human says stop it right now.

I used Claude Code to write a relatively complicated watchOS app. I know how to program (FAANG L5), but didn't really know Swift. I achieved a pretty good result for about $600, while a contractor would've cost much more.

It's almost vindication for where I work an SDE needs to do everything, infra, development, deployment, launch, operations. There's no dedicated QA, test or operations on a product level, and while AI helped a great deal it's pretty clear it cannot replace me at least within the next 2 to 3 iterations.

If I was only writing code, the fear would be completely justified.

I've been using Claude Code + Opus for side projects. The only thing that's changed for me dev wise is that I QA more, and think more about how to solve my problems.

I've mainly been using Sonnet 4.5 so decided to give Opus 4.5 a whirl to see if could solve an annoying task I've been working on that Sonnet 4.5 absolutely fails on. Just started with "Are you familiar with <task> and can you help me?" and so far the response has been a resounding:

> Taking longer than usual. Trying again shortly (attempt 1 of 10)

> ...

> Taking longer than usual. Trying again shortly (attempt 10 of 10)

> Due to unexpected capacity constraints, Claude is unable to respond to your message. Please try again soon.

I guess I'll have to wait until later to feel the fear...

It definitely feels like a jump in capability. I've found that the long term quality of the codebase doesn't take nosedive nearly as quickly as earlier agentic models. If anything it's about steady or maybe even increasing if you prompt it correctly and ask for "cleanup PRs"

Ironically AI may replace SWE way faster than it does for any other businesses in Stone Age.

Pick anything else you have a far better likelihood to fall back into manual process, legal wall, or whatever that AI cannot replace easily.

Good job boys and girls. You will be remembered.

I agree with this guy: https://obie.medium.com/what-happens-when-the-coding-becomes.... It's a game changer when used to pair program (MOB Programming) with you as the navigator and the LLM as the driver. It needs guidance and rework but shit gets done with the human as the ultimate gatekeeper.

This just looks like an advertisement?

flagged for astroturfing

Reads like astroturf to me.

> do not know what's coming for us in the next 2-3 years, hell, even next year might be the final turning point already.

What is this based on? Research? Data? Gut feeling?

> but how long will it be until even that is not needed anymore?

You just answered that. 2 to 3 years, hell, even next year, maybe.

> it also saddens me knowing where all of this is heading.

If you know where this is heading why are you not investing everything you have in these companies? Isn't that the obvious conclusion instead of wringing your hands over the loss of a coding job?

It invents a problem, provides a time line, immediately questions itself, and then confidently prognosticates without any effort to explain the information used to arrive at this conclusion.

What am I supposed to take from this? Other than that people are generally irrational when contemplating the future?

Not me, but it's the first that for now didn't fail catastrophically on moderately difficult task. Not that the other models can't manage difficult tasks, but they sometimes generate stuff so wrong it's almost funny. For now (a week of usage), it never truly failed (the one time it couldn't find a working solution, it simply proposed solutions I could use and stopped, without generating bullshit code), and sometimes I don't have to change anything from the generated output.

It's definitely more useful than me the first 5 years of my professional career though, so for people who don't improve fast or for average new grades, this can be a problem.

Same here. Using it this week and on Thursday I began to understand why Lee Sedol retired not long after being defeated by AlphaGo. For the stuff I'm good at, 3 months ago I was better than the models. Today, I'm not sure.

I wonder if these coding models will be sustainable/profitable long term if the local models continue to improve.

qwen3-coder blew me away.

Honeslty I have a lot of friends who are studying SWE and they are saying the same thing do you guys think that if they do get replaced they'll still be needed to maintin the AI's.

I tried it and I'm not impressed.

In threads where I see an example of what the author is impressed by, I'm usually not impressed. So when I see something like this, where the author doesn't give any examples, I also assume Claude did something unimpressive.

That’s only tangentially related but I have a very hard time using Opus for anything serious. Sonnet is still much more useful to me thanks to the context window size. By the moment Opus actually understands what’s needed, I’m n compactions deep and pretty much hoping for the best.

That’s a reason why I can’t believe the benchmarks and why I also believe open source models (claiming 200 but realistically struggling past 40k) aren’t only a bit but very far behind SOTA in actual software dev.

This is not true for all software, but there are types of systems or environments where it’s abundantly clear that Opus (or anything with a sub 1m window) won’t cut it, unless it has a very efficient agentic system to help.

I’m not talking about dumping an entire code base in the context, I’m talking about clear specs, some code, library guidelines, and a few elements to allow the LLM to be better than a glorified autocomplete that lives in an electron fork.

Sonnet still wins easily.

OpenAI is burning through $60B a year in losses.

Something doesn't square about this picture: either this is the best thing since sliced bread and it should be wildly profitable, or ... it's not, and it's losing a lot of money because they know there isn't a market at a breakeven price.

> Sure, I can watch Opus do my work all day long and make sure to intervene if it fucks up here and there, but how long will it be until even that is not needed anymore?

Right: if you expect your job as a software developer to be effectively the same shape on a year or two you're in for a bad time.

But humans can adapt! Your goal should be to evolve with the tools that are available. In a couple of years time you should be able to produce significantly more, better code, solving more ambitious profiles and making you more valuable as a software professional.

That's how careers have always progressed: I'm a better, faster developer today than I was two years ago.

I'll worry for my career when I meet a company that has a software roadmap that they can feasibly complete.

Not sure I’d be worried for my job, but it’s legitimately a significant jump in capabilities, even if other models attempt to fudge higher bench results

"This model is so alive I want to donate a kidney to it!"

Opus 4.5 is like a couple points higher then Sonnet 4.5 on the SWE benchmark.

How does Opus 4.5 compare to gpt-5.1-codex-max?