The Four Tiers of AI Intelligence

There are many ways to group AI or LLM intelligence.

In this one I weakly categorize them based on the difficulty of getting there, as influenced by their relative position to other LLMs.

Also note, due to the pace of AI development, a model is in a tier at the point of release (+/- a few days). You may fall behind after and most likely will.

Lab specific analysis are as of today, November 10th, 2026. If you’re coming here later they’re probably outdated.

The rule is: it is much easier to move within a tier than across them. Some of them are quite broad, so I’m not saying crossing the entirety one tier is harder than moving to the next tier, but rather moving up a few points on benchmarks when it crosses a tier is much harder than those same points within one. The reasons for each are explained in their respective definitions.

To get here you need to have some of the best people in the world, tons of high quality data and an absolute glut of compute. You are also past the point where you can rely on any type of distillation or learning from your competitors because you are better than them. You have to teach your model how to get here by directly iterating against the real world, no tricks.

The only labs here are the big three: Anthropic, OpenAI and Google.

Other labs? I don’t see a good chance for any right now. Muse maybe, if they manage to improve on their Wonder Full line.

This one is easier to get to, but still very hard. The main reason, apart from just not needing to be as good, is you have access to the T1 models, both to do the R&D, and as a source of data and research. Access is limited of course, all three labs in T1 ban using their models to build a competitor, with Anthropic taking the most aggressive steps to prevent it after becoming the distillation target of choice for many labs. You don’t have access to full reasoning traces or weights or even model architectural decisions from the T1s.

Those are only available from open source models, and like in T1, you can’t rely on those because you are better than them.

Reaching T3 is certainly not easy but it is a lot more achievable. You now have access to the top open models to learn from and do R&D for you, with their full reasoning traces, their weights, their full architecture. Of course, you still need very smart people, lots of compute, data etc.

This is still a good place to be, as long as you have some other advantage such as cost or speed.

This is the most crowded one, so I’ll list more briefly:

I considered making another tier like “Still good at current tasks” for more separation or to clarify just because it is here does not mean it is bad (though many models here are bad). It may still be an excellent model if it does well in something like size or cost.

As an example take Qwen 3.x 27B. Not the strongest model but still quite strong and among the most loved, because it packs all its power into just 27 billion parameters, small enough to run locally, if you can afford a good gpu or a high end mac.

Good models here include Qwen’s local models, Step 3.x, Gemini Flash-lite, Mimo flash, Deepseek V4 flash, GPT Luna.

Many models here are still fine and quite interesting, they’re just not competitive with their peers. This includes Mistral, Ring/Ling, IFM’s K2 horizon, Beam, Nemotron, Minimax M3.1 flash, Trinity, Mercury

This article may miss a model or lab in in T2 & T3, so let me know in the comments of wherever this is posted.