For lots of use cases you don't want the BEST model, you want the model that gets the job done for as little money as possible.

I disagree, I always reach for the best model when I’m at work. For normal questions I do use the default ChatGPT model.

Fundamentally, people underestimate the marginal increase in utility for marginal increase in “intelligence” as measured in benchmarks. Like I’d pay 5x for a model that has 2 units higher in AA intelligence.

You could tell me AA int scale is logarithmic and I would believe you

And that's almost always another OpenAI or Anthropic model, which are cheaper and more reasoning-efficient

> Claude has a different personality and performance than other models for example

This kind of thing amounts to a LoRa applied to your model. There's no reason (apart from potential IP infringement) that someone else couldn't produce a "Claude LoRa" for their model (or, more likely, given the dislike for Claude's writing style, Anthropic could offer a "non-Claude LoRa" for their own models)

I completely disagree with this. What you are saying is that performance, all they need a simple trick and they get Claude performance for free. In that case, why do we see open weight models still lagging?

Assuming you meant this for personality, I’d still disagree but not that extensively. I think personality is nuanced and complex - it can’t be simply lifted and shifted. It’s grown using a meticulous process.

> I completely disagree with this. What you are saying is that performance, all they need a simple trick and they get Claude performance for free. In that case, why do we see open weight models still lagging?

Yes, they are lagging.

But "performance of Claude's from 6 months behind" is fine for a lot of tasks. And the performance of today's Claude will be still useful for many (and more and more) tasks in 6 months.

The frontier labs are making models for smaller and smaller share of the pie, as the capabilities of open models grow. That isn't sustainable

> But "performance of Claude's from 6 months behind" is fine for a lot of tasks. And the performance of today's Claude will be still useful for many (and more and more) tasks in 6 months.

Okay, then why would one buy 6-months-ago performance for 2-3x more from an open weights lab as opposed to a closed lab?

OpenAI and Anthropic are competing heavily with GPT-6 Luna and Claude Haiku 5.5!

Claude’s personality is the worst aspect of the product!

Disagree but I don’t think this negates my point

"Cursor" Grok has the exact same (non-)personality as Claude.

GPT has the same performance. And performance is definitely a commodity. Why wouldn't it be?

I don’t think GPT Astra is as good as Opus 5.5, it’s very noticeable how pedantic and poor at intuiting things it is to me. It often calls things done when Claude would keep working on them, and doesn’t seem to have as good of an eye - often leaving clear visual defects and calling things done.