Hacker News new | past | comments | ask | show | jobs | submit
It seems like everyone I know outside of tech uses ChatGPT “Chat” on the free tier.

Of the people I know in tech, other than those who don’t use either, everyone is paying for Claude.

It's pretty obvious that OpenAI is dieselgating* the benchmarks. Even 6.2 Astra is nowhere near as good as Claude models for coding. Any software engineer can do a side-by-side comparison (as many of us do) and figure it out.

*Also consider the possibility that Anthropic may be dieselgating it down ("benchminning"?) instead of OpenAI doing benchmaxxing. It might make sense from some game theory standpoint - If you are significantly better than your enemy, it might be in your interest to lull them into believing that you are only slightly better.

I do coding evals every time a new model comes out (that I have access to) and Opus 5.5 is a touch better than Sol 6.2 right now but that’s a recent development. Prior to the 5.5 release open ai outperformed at every tier.

Sol is half the price of Opus and Luna outperforms haiku at every eval I throw at it. Haiku wasn’t even able to complete several of my evals until the newest release.

The idea that Anthropic is miles ahead of OpenAI doesn’t pass the smell test in my usage.

And that’s before the Claude vs Codex comparison (which I’ll grant is more personal preference but I find Claude to be wretched).

> Also consider the possibility that Anthropic may be dieselgating it down ("benchminning"?) instead of OpenAI doing benchmaxxing. It might make sense from some game theory standpoint - If you are significantly better than your enemy, it might be in your interest to lull them into believing that you are only slightly better.

Unlikely, because it would mean that you're also "lulling" your customers and your investors into believing that you're only slightly better than your competitor ("enemy" isn't the right term here), and you don't want that. Especially not shortly before your IPO.

That's not obvious at all to me. While I prefer Opus for run-of-the-mill tasks I actually find Astra to be better for very hard problems. It's just a bit of a psychopath and has more "paperclip maximizer" tendencies. If my hypothesis is true then how they fare head-to-head depends on the workload.
> Any software engineer can do a side-by-side comparison (as many of us do) and figure it out.

That's a really strange take to me, because among my colleagues theirs basically a split between two kinds of people:

- the “I don't need frontier so I use sonnet and it's good enough” group.

- and the “I want the best model and I managed to get a Codex subscription from management because Sol is much better than Opus” team. (This only started this summer and I suspect it was mostly true because of Opus 5, but that's still going to do long term damage).

No one is arguing the other way around and I know the manager would be considering a switch of what comes as default to full OpenAI on the first occasion (he's a very cost sensitive kind of guy so I'm sure that any price hike would trigger him).

Codex doesn't have Claudish, though, and you can use other harnesses (i.e. Pi with its immense tokens savings) with a coding plan.

I use Codex for lack of Claudish _alone._ My spidai senses might still be overtuned, but I can spot Claude generated content a mile away and my brain involuntarily turns off.

Also, the free chat in ChatGPT is so annoying that I don't even want to try their coding agents.
Pretty much. I keep urging my non-tech friends to pay for Claude and they keep using free tier "Chat" like a fancy search engine and copywriter
Well, why would they do that? What use could they get from paying for a somewhat better search engine and copywriter?
Running claude on your personal computer as a non-technical user is reckless bordering on insane. You have absolutely no idea what it's doing!
I’m curious, what’s your arguments?
Do you own shares in Anthropic or is there any reason why you urge people to buy something they don't need?