Claude Haiku 4.5 DESTROYS GPT-5?

A faster and cheaper model that is great at coding is here. It's Haiku 4.5 from Claude and it might become my default model for all my coding needs. It genuinely has. I've been using this all week. Well, I say all week. It launched pretty much yesterday. I've been using it all day yesterday and it's pretty it's really really faking fast and it's better than Sonnet was like five months ago. Claude have set, they have haiku and what was the name of opus. So the idea is that the opus is the frontier model like the best and most expensive model right seemingly theoretically. Um and then we have set which is like the medium range model and then we have haiku that is the light fast cheap model but now the light fast cheap model is better than the other two. The light fast and cheap model is better than what sonic was 6 months ago 5 months ago. So, Haiku now is 33.3% on Sweetbench verified whereas Sonet 4, which is what we were using dayto-day was is 72.7. So, it's better than what Sonic 4 is. Sonet 4.5 has since been released, which is now better than Opus and previous Sonic versions. So, five months ago, Claude Sonic 4 was state-of-the-art. Today, Haiku 4.5 matches its coding performance at onethird of the cost. But the interesting thing is here it's not as good as GBT5 codecs. Um it's better than GBT5 on the Sweetbench verified and much better than Gemini 2.5 Pro. It the these are comparable to Frontier models whereas this is their fast and cheap model. They're not comparing it to realistically this should compare it to GPT5 Mini or Gemini 2.5 Flash but it's not. It's comparing it to the other company's Frontier models. So, it's it's a really really good model to have. Now, it has increased in price. It's $1 in, $5 out, where it used to be 8.7 in and then $3 out or something like that or $4 out. So, it is, you know, um I want to go backwards actually. It is more expensive than what it was before, but it is a really good model. I have put all of my sub aents as haiku. So sub agents, if you don't know what they are, they're kind of specialists that run in their own context window. So I have a TypeScript specialist that goes and fixes all my TypeScript errors. They don't need a lot of they get told what to do. They don't need a lot of thinking power. They just need to churn out whatever it is they're good at. And yeah, so and because of that speed as well, not a lot of people have it, but you've got the the Clawed Chrome uh plugin, so it's great. It's just great at like quick summarization data. It's just a just a really freaking good model to be fair for those sorts of tasks. Now what I want is I want to be able to um what they used to have before Sonnet just became the best is you used to have Opus plan mode that would degrade to Sonnet degrade to Sonnet uh whenever it is you implemented stuff because you that's what you kind of want. You want the big boys to do all the thinking, the planning, the heavy grunt work, and then you want to delegate to the the runs, you know, the guys who can just don't need who just do the leg work. Do the leg work, right? And that's what I loved. I would do the plan mode, Opus, and then go to Sonnet. They've removed that since just became the de facto. Um, I want that back now, right? I want I want plan mode, Sonnet, and then Haiku. And then when I do the work, it goes to haiku. So we can just churn it out and not use a lot of um tokens and whatever. So what I've done, I've built my own slash command called plan. So when I go slash plan, you can choose your model that that plan mode runs in. So I leave I leave my default model as haiku slash plan do this that and the other and that gets set as sonnet. Then that's sonet 4.5. So it will run that and then when I actually do the work or when I don't run the plan work, sorry, when I don't run plan command. So, I've kind of rebuilt that thing. Um, that that system that I used. Um, but yeah, it's it's it's just great. It's it it just it's just a testament to how far we've come really, you know, in just five months, what was the groundbreaking frontier model, state-of-the-art model is now being beaten by little cheap thing, you know. Yeah. Saying all that, I I've canled my claw code subscription. Um, I'm going to commit to GLM 4.6. six next month. Um, but this definitely made it slightly more tempting to stay on cord. This is part of a larger conversation on my show, Command AI, which we stream live every single week. We discuss the news and all things related to AI in the world of design and web. Catch us next week and join in the banter.