Hacker News Reader

Comment by lnenad | Hacker News Reader

Every single model I've tested that is supposed to be anywhere near Opus, wasn't there. There is some magic sauce in Anthropic models that these benchmarks aren't demonstrating.

View on HN

It's been suggested elsewhere that the AA-Omniscience benchmark [1] is more representative than the rest. It puts K3.1 about 6 months behind the frontier and the new version of DS 4.1-flash about 12 months. That roughly lines up with my gut feeling as well, at least when it comes to the trickier problems. Still rather impressive, but not quite frontier.

1. https://artificialanalysis.ai/evaluations/omniscience

+1 for the omniscience benchmark!

Although even this one needs to be taken with a grain of salt. Fable/Opus still act very differently to Astra and Argon imo. But maybe it's just because I am used to these. I wouldn't clearly put them above each other though, although I still prefer anthropics models, but that could be simply out of comfort.

Huh, this is an interesting hypothesis! It would make sense, to a degree. Part of "big model smell" to me is the model being able to pull a huge range of interesting algorithms, architectures, mathematics, and everything else that would come with greater world knowledge and having a much more sort of innate, like, full, well-rounded understanding of it instead of surface-level ICL pattern matching on web results.

Yeah that tracks completely with my experience as well.

That definitely matches my experience, I have found Astra to be more capable than many other benchmarks show, have been very impressed with Argon for text, and generally like Sol. I find Haiku and Sonnet to be not worth my time as compared to the speed and cheapness of something like DeepSeek.

The only exception on the chart is DeepSeek - it scores exceptionally low there, which doesn't exactly track my experience. It usually does what you tell it to, and it's not exceptionally intelligent, but I do find it quite good at what it does (the bulk of code, scripting, etc). I know a lot of folks like it too ... curious if I'm just not exploring the simpler models enough?

I skimmed through the publicly available questions. The first programming-related one I saw has the answer in the question, not a great start :D [1]

  Q: In C23, attributes are allowed in an unbraced (forward) union declaration only in the form “union [[...]] Tag;”. What single character must immediately follow the identifier in that declaration for the attributes to be permitted?
  A: ;

[1] https://huggingface.co/datasets/ArtificialAnalysis/AA-Omnisc...

Edit: The other programming ones look fine, bad luck I saw this one first I guess.

Maybe I need to try it again, but I have never liked the opus models.

I keep having problems with them not doing the thing I ask, (and also failing to call tools correctly in pi. I think they've very much trained to the harness, which is frustrating.).

I always found them too chatty and verbose. I'd get a wall of text and miss things because it was lost in garbage about load bearing and blast radius

I think this part has been prettymuch fixed in recent releases. (I didn't just get used to it, because when I look back at older output, it still drives me crazy with all of these buzzword tics).

I recently just crossed the threshold for 100% decompilation of a PlayStation game. I used a lot of models in the process, and I feel like this is a good demonstration the "sauce" that opus has.

The harness I've built for it is pretty straight forward: split binary into individual functions, give the agent a bunch of small tools to diff output with underlying binary, make them wrote clean c until there's a certified match.

Rin a final oracle to make sure the agent isn't cheating.

- Qwen 3.8 27b on my 5090, barely acceptable, thinks too long, sometimes freaks out.

- deepseek flash variants, work horse. Runinfra provides around 300 tok/s. Great for tearing through moderately complex functions

- glm and kimi k.3, expensive, slow, ultimately not that much better than deepseek flash for this task. Better at doing long context research into old compiler productions

- sol 6.1, best bang for buck. Creative enough to methodically work through more complex spellings

But I found there were around 1% of functions that these models simply cannot solve. Opus and Astra take them handily.

I'm still still in the process of looking through the solution traces, but from watching the missing sauce seems to be creative reasoning.

Sol 6.1 will just happily grind on permutations of a base shape spelling that M2C outputs, where opus and Astra immediately realize that they need to unwind assumptions taken for granted (and sometimes in the process endure a worst match).

I'm hoping to be able to capture that creativity step and distill it into the harness tools.

A lot of these models are likely benchmaxxing the existing benchmarks that are out there. Some perhaps even by accident.

I think looking at some more niche benchmarks helps, I created one and it does show Kimi K3.1 Pro as the most intelligent of all open-weight models, though it is still significantly below Opus 5/5.5[1]

1 - https://bench.killswitch-lang.org/

This seems more a measure of anti promp injection than general intelligence

I'd say it's both. Models should be intelligent enough to be able to tell prompt injections apart. But the language itself is also novel and not in the model training data set, so that should be a good signal for intelligence in itself.

It really depends on the type of task you're doing and how you're using it, because for the past two month the only reason I've had to use my company's paid Opus plan instead of my local Qwen-3.8-Next was speed. I use Qwen as a back-up plan when my token limit is reached and in my day to day work they have been indiscernible in terms of output.

I agree, but my point wasn't about being able to use a weaker model, it's about benchmarks not being in any way representative of the actual quality/level of the model. Of course for a lot of scenarios Opus is overkill and Qwen/DS are more than enough.

I think also heavily depends on what you're doing.

If you're in a stable codebase with lots of best practices, patterns, controls (linters, analyzers, etc.), examples already established, then any model will do ~the same plus/minus since there's not much novelty.

If you're building ground up, entirely novel, no reference, I think results will vary with lower capability models.

For me, I do a mix, leaving implementation to Haiku/Luna/DeepSeek and research/feedback/critique to Opus/Sol/Astra/Fable

I think the secret sauce is about communication. It understands you and you understand it. Just a better coworker to work with. Benchmarks don’t measure that.

Yeah, but if you need to provide less details to Opus to get the same output it means the model is better at deducing your needs, which again implies it's a higher quality model.

Opus 5.5 is flat out amazing, particularly how cheap it is in practice (I'm on a subscription and use Claude Code and I almost never even approach the cap).

I would love a benchmark suite that better reflects real-world coding performance, as having Muse Wonder in the top three makes no sense to me.

I spent the last 4 months on a mix of chatgpt and glm building some prototypes of product ideas I've had for years because claude was in the phase of obnoxious conversation and verbosity.

I took the output and original specs and handed them to Claude today and the level of insight and concise detail in it's criticism and problem solving was fantastic. Dropped the others immediately

Same. Every time I prompt some other new/best/benchmaxxed model, I think to myself: wow, finally, this is amazing. And then I give the same thing to Opus and the bastard is always better. It's annoying to a point.

Did you try it the other way around and ask GLM to review opus? Otherwise you didnt learn anything.

It's not just 5.5, every iteration of the frontier moving - Opus has had some edge. I'm sure it's not 100% in every scenario, I love interacting with Codex models more, but the quality of the output, in most scenarios that I've been exposed to, has been the best.

Anthropic cooked with Opus 4.5 and 5.5. Their other iterations have usually been nicer to interact with than GPT and more "autonomous", while being dumber.

Can confirm. Labs doing benchmaxxxxing and struggle to stay relevant

Reply on news.ycombinator.com