Startup claims it solved a long-standing AI problem

Subquadratic, a Miami startup that came out of stealth in May 2026, says its SubQ model solves the quadratic attention bottleneck that has limited large language models for almost a decade. According to MIT Technology Review, the performance gains are dramatic: up to 56 times faster than models using FlashAttention, a context size of up to 12 million tokens, and 98% accuracy in needle-in-a-haystack retrieval tasks involving 6–12 million tokens of context.

This bottleneck is real and longstanding. In a typical transformer architecture, “attention” is “dense”: each token attends to all other tokens, meaning that doubling the length of input quadruples the amount of computation required. This makes working with long texts computationally expensive, and restricts how much context we can use. SubQ replaces dense attention with a “sparse,” text-dependent mechanism that only pays attention to relevant tokens. Founded by CTO Alex Whedon and CEO Justin Dangel, the company argues that this allows them to achieve both speed and scalability.

1The numbers — and the asterisks

Their cost claims are especially bold: according to the company, they were able to run the RULER 128 benchmark for just $8, compared to $2,600 they say was spent by Anthropic on their Opus 4.6 model. They also achieved 89.7% on the LiveCodeBench competitive coding benchmark, and reportedly completed a 400-document analysis job in just a few seconds. Their work was evaluated by third party Appen, who described the approach as having “potential to be a game changer.”

But wait… There’s fine print. Access is currently limited (there’s a long waitlist, even though they claim to have tens of thousands of people signed up), so very few people have had a chance to really push the model. SubQ started with pre-trained weights from the Chinese open source Qwen model, rather than training their own weights, making it hard to directly compare. And AI researcher Will Depue summarized the doubters’ perspective succinctly: “the current public evidence does not yet justify the stronger claim that they have solved the quadratic attention bottleneck.” Remember that benchmarks are not the same as generalization to diverse real world applications.

Even if you don’t use SubQ, why should you care? Because the quadratic barrier is why your current models struggle to remember what you said at the beginning of a long conversation, or understand an entire codebase, or pay a significant fee to read a long document. If we can solve this problem efficiently, many limitations will fade away: you’ll be able to feed a model an entire book, a year’s worth of email, or a huge code repository, and expect it to keep all the information in its head. I think the hype is real, and so are the concerns, and we should be excited but cautious until we see others try it out. If this is truly a breakthrough, it will be huge — which is why we need to demand better evidence than some initial benchmark results.