The Gap Is Three Points

Six months ago the consensus was that AI was about to split into haves and have-nots.

The reasoning was clean enough. Frontier models cost billions to train. A handful of companies can afford that. Those companies price access accordingly, everybody else rents scraps, and if you built a business on top of somebody's API you built it on rented land with a landlord who sets the rent.

I want to walk through why that didn't happen, because the evidence is now hard to argue with, and it's better than I expected.

The number that matters

Artificial Analysis runs an independent Intelligence Index across a few hundred models. They aren't a lab and they aren't selling a model, which is exactly why I use them instead of vendor benchmarks.

As of August 2026 the top of that index is Claude Opus 5 at 63.

The best OPEN WEIGHTS model on the same index is Kimi K3 at 60.

Three points is the whole distance between the most capable model on the market and one you can download, run, fine-tune and ship without asking ANYONE's permission (and yes, you can literally just download it). Downloading is the easy part. Kimi K3 is 2.8 trillion parameters, so in practice you're running it through a host rather than on your own hardware. But the weights are out, and nobody can pull them back. Behind Kimi you get Qwen3.8 at 58 and DeepSeek V4 Pro at 53. Of the roughly 180 models they track, more than 80 are open weights.

Here's the part that got my attention. A couple of weeks ago that second-place open model was sitting several points further back. The gap isn't just small, it's closing in real time.

Two years ago you gave up a lot to run open. Today, for most production work, you're giving up almost nothing.

What actually happened

Three things, and together they look like a pattern.

1. Meta opened up its weights. It released a dense 30-billion-parameter model called Muse Glimmer that runs on a laptop with a single consumer GPU, and said a version of its most capable model would follow. Zuckerberg published a 6,500-word essay called "The Future is for Everyone" making the case that the real risk in AI is one entity holding too much control. Agree with him or not about motive, the artifact is the weights and the weights are out.
2. Grok 4.6 is priced like it wants the volume. $2 per million input tokens and $6 per million output at normal prompt sizes, with a 500k context window. Push a single request past 200k tokens and the whole request reprices to $4 and $12, so check the current page before you budget against it. These numbers move. You're getting frontier-adjacent capability at commodity pricing.
3. The Chinese labs are setting the pace on open weights. Kimi K3, Qwen3.8 and DeepSeek V4 Pro hold the top three open slots. Those are Moonshot, Alibaba and DeepSeek, which is exactly the competition Zuckerberg pointed at when he warned that Chinese developers could take the lead if the US makes open AI harder to build at home. When your rival's public strategy memo names the pressure, the pressure is real.

Why a producer cares about any of this

I run production pipelines. Every pipeline is a chain of steps, and every step has a supplier. Storyboard generation. Image generation. Upscaling. Video. Voice. Cleanup. Each one is somebody's API with somebody's pricing page and somebody's terms.

The thing that keeps producers up at night is supplier concentration. One vendor per step means one phone call can reprice your entire show, and you find out in the middle of a delivery week. I've watched that dynamic play out before with render farms and stock libraries, and it goes the same way every time.

What changed is that there are now four or five credible answers to nearly every step in that chain, and a real fraction of them are weights you can hold. If a vendor triples their price on a Tuesday, you're looking at an afternoon of work instead of a renegotiation you lose.

You now have a structural protection on your cost basis, and you got it because a handful of very large companies decided to compete instead of settle into a cartel.

Practical version, three rules I actually run by:

1. No step in the pipeline gets exactly one supplier. Every step needs a named alternate you've actually tested, not one you assume would work.
2. Keep at least one open-weights option live in the chain. It doesn't have to be your default. It's your insurance for the week a vendor changes the terms mid-project.
3. Re-price the whole pipeline quarterly. The floor moves fast enough now that last quarter's rational choice is this quarter's overpayment.

The part nobody puts on the slide

Cheaper per token doesn't mean a smaller invoice.

Usage climbs to fill whatever the price allows. I've watched teams cut their per-token cost dramatically and still write a bigger check at the end of the month, because the cheap tokens invited five new workflows that didn't exist before. Often that's a good trade, and it's still not a saving.

The number worth tracking is cost per finished deliverable. One approved 30 second spot. One delivered episode. One shipped campaign. The per-million-token figure will tell you a flattering story that has very little to do with your budget.

I'd also hold the celebration a little loosely. Open weights today is a competitive strategy. Strategies get revisited. Meta opened up because being second in a closed race is worse than being first in an open one, and if that math changes the policy can change with it. Build like the floor is real, because right now it is, and stay awake to the fact that somebody chose to put it there.

The floor came up under everybody. Excited to see what gets built on it.