Picking a model by hype is expensive. The signal that actually matters for business work is cost per task — how much you pay to get a real job done, not how many tokens or benchmark points a model burns. Watching that single column changes which model you reach for.
Because you have no incentive to like intelligence for its own sake. You want to keep your costs down while the work still gets done, and the frontier charts make that trade-off visible at a glance.
Reading the Pareto frontier for business work
On the cost-per-task frontier, a few models stand out. GLM 5.3 Flash was really good for a while — then Haiku 5.5 came along and smoked it, delivering more intelligence for a cheaper cost. Mimo 2.6 Pro also sits near the top: a really fantastic model for very, very cheap. OpenAI is dominating the frontier too with GPT 6.1 Soul, which is surprisingly good for how cheap it is.
But for the majority of business tasks — reading email, summarizing, booking calendar slots, coordination work — you want to sit in the cheap-and-good-enough zone, not at the top of the capability scale. You only need the expensive end if you're doing scientific research or trying to solve all of math and physics. For everyday business work, the green space of the chart is where you want to be.
Classifier work is a different game
The second thing to watch is which model type fits the job. There are two use cases now: classifier-style decisions and unstructured-to-structured work. For classification, you hand a model some text and ask a yes-or-no question — is this about Jira, should we turn this into a social media post? — and pipe the answer along.
Jev got a lot of hype, but Liquid AI's D1 Ultra already outclasses it significantly — and that's the one to use now.
That's what makes D1 Ultra from Liquid AI the new pick for this kind of work. Jev got a lot of hype, but these classifier models are already outclassed significantly by D1 Ultra — so if the job is making quick decisions and passing the result downstream, that's the model to be using now.
Keep checking back as the frontier moves
The frontier updates regularly, and a new DeepSeek release just landed that is apparently very good and will show up on the chart soon. The habit that pays is the same every time: look at cost per task, match the model to the job type, and ignore the hype cycle. That's how you know what you should actually be using to get the job done.
Transcript
All right, so here's what you should be paying attention to in the AI space or how to get a quick glance of what's going on. Uh, so you want to be paying attention to cost per task. Uh, that's very important for your business and for the work you're doing, you have no incentive to like intelligence and token max. You want to keep your costs down and so you can see how the models are performing here. Uh, you know, you got glm 5.3 flash, which is, was really good. And then haiku 5.5 came uh, along and smoked it. More intelligence for cheaper cost. Um, and then of course Mimo, uh, two uh, point six Pro is up here. Really fantastic model, uh, for very, very cheap. Um, so that's essentially the information you ah, wanted. This is the Pareto, uh, frontier. You can see, uh, OpenAI is actually dominating it with GP, ah, ah, GBT 6.1 soul. I, I'm really surprised, um, how good um, and cheap that model is. That's, that's wild. Um, but Mimo is smoking them. Um, and really where for the majority of tasks in business it's, you want to be up here because it's just coordination tasks like reading, email, summarizing, booking a calendar. You just don't need to be up here unless you're doing like scientific research and trying to solve all of math and physics. Um, but for the majority of uh, business work, you want to be, uh, you want to be in this green space and you really want to be on that guy. Now the next thing to you want to know about is you want to know about, oh, where's the decision? This one, this was actually new information for me today, this D1 Ultra. Because, uh, Jeff, these models are classifiers. You basically give it some test text and you ask, oh, is this talking about, uh, you know, Jira or is it talking about, uh, you know, whatever. That's what I'd use it for. Uh, should we make a social media post out of this? Whatever. It just makes a decision and gives you back yes or no and then you pipe that data along. So Jev got a lot of hype, but you could see they're already outclassed significantly by these liquid AI guys. D1 Ultra. Um, so it's like, oh, actually this need to do this and so I'm going to be using this now. So I thought I'd share this resource. It's very important if you want to know what you should be using, um, to get the job done. Um, it should be, it should be this. And I think there's a new Deep Seek today that was released that's going to be coming up on here. So we'll see how it's performing. Apparently, it's very good. Um, but, yeah, this is this updates regularly. All right, thanks for watching, guys. Take it easy.
Frequently Asked Questions
It is the amount you pay to get a real task completed, and it is the column to focus on because you have no incentive to pay for raw intelligence or token maximums — you want costs down while the work still gets done.
The majority of business work is coordination tasks like reading email, summarizing and booking calendar slots. You only need top-end capability for things like scientific research or solving math and physics.
It is Liquid AI's D1 Ultra. Jev got a lot of hype, but the classifier models are already outclassed significantly, so D1 Ultra is the one to use now.
Classifier work means giving a model text and a yes-or-no question — for example, is this about Jira, or should we make a social media post out of this — then piping the answer along.