So there’s this company called Subquadratic that came out of nowhere on May 5, 2026, and the AI world kind of lost its mind for a few days. Their model, called SubQ, got 12 million views on X within 24 hours. Thirty thousand people joined the waitlist before most of the tech press had even written about them. And the reaction was split pretty cleanly down the middle half the people saying this is the biggest thing since ChatGPT, the other half calling it vaporware from a company nobody’s heard of.
I went through everything. The launch posts, the benchmark numbers, the skeptics on Hacker News, the technical blog they pushed out within hours of the first wave of pushback. And I want to explain what they’re actually claiming, why it matters if it’s true, and why I’m not fully convinced yet, but also not dismissing it.
Let me start from the beginning, because the problem they’re trying to solve is worth understanding properly.
The Math Problem at the Heart of Every AI Model
Every AI model you use today, ChatGPT, Claude, Gemini, all of them: is built on something called a transformer architecture. The “T” in ChatGPT literally stands for transformer. This architecture was introduced in a 2017 paper from Google and it basically took over the whole field. It’s how these models understand language so well.
But here’s the issue with transformers, and it’s a math issue. The way they work is that every word (or “token”) in your input has to compare itself to every other word. So if you send 100 words to an AI, it does 100 × 100 = 10,000 comparisons. Send 1,000 words, it does 1,000 × 1,000 = 1 million comparisons. Double the input, and the work doesn’t just double. It quadruples. That’s what “quadratic scaling” means — the compute explodes as the input gets longer.
This is fine when inputs are short. But try to feed something really long, like a full legal contract, or an entire codebase, or a 400-page report and costs go through the roof. Most AI companies just cap the input length and move on. The ones who don’t cap it charge a lot for longer inputs.
The workaround that most people use is called RAG, which stands for Retrieval-Augmented Generation. The idea is pretty simple: instead of feeding the whole document to the AI, you cut it into small chunks, store those chunks in a database, and when a question comes in, you search for the most relevant chunks and feed only those to the AI. Think of it like tearing a 400-page textbook into individual pages and keeping them in a box. When someone asks a question, you rummage through the box, pull out the five pages that seem most relevant, and show those to the AI.
It works. But it’s messy and it misses things. The connection between chapter 2 and chapter 17? Gone. A detail in the appendix that contradicts something in the main body? Your AI probably won’t catch it. RAG exists only because of this one math constraint. Subquadratic says they’ve fixed the constraint itself.
What SubQ Actually Does
Their CTO is Alexander Whedon, who was at Meta and then ran generative AI at TribeAI. The team has PhD researchers from Meta, Google, Oxford, and Cambridge, 11 of the 13 people in the company are PhDs, which is an unusual ratio even for an AI startup.
They built something they’re calling Subquadratic Sparse Attention, or SSA. The core idea is this: instead of making every token compare itself to every other token, the model figures out which comparisons actually matter and skips the rest.
Imagine a big company meeting where the standard rule is everyone has to introduce themselves to everyone else. With 100 people in the room, that’s 4,950 introductions. SSA says: no, the marketing person only needs to talk to other marketing people and maybe the sales team. The engineer doesn’t need to shake hands with the accountant. You go from thousands of conversations to a few hundred, and you don’t lose anything important in the process.
The result is a model with a 12 million token context window. To give that some scale: one token is roughly 0.75 words, so 12 million tokens is around 9 million words. That’s something like 120 full novels in a single prompt. Their CEO Justin Dangel described the goal simply: “We are trying to figure out how to not compare every token to every token to every token.”
They launched three products simultaneously on May 5, all in private beta right now. First, an API that exposes a 1 million token production model through endpoints compatible with OpenAI’s format, which means developers can swap it in without rewriting their code. Second, the 12 million token research model, which is gated to select partners for now. Third, a coding agent called SubQ Code, which loads your entire codebase into context and reasons across all of it at once. There’s also a fourth product called SubQ Search, but details on that one are still thin.
They raised $29 million in seed funding. For a 13-person company, that’s a serious amount.
The Benchmark Numbers
This is where things get interesting. And also where the skeptics have the most to say.
On a long-context benchmark called RULER 128K, which tests whether a model can actually find and use information spread across a very long document, SubQ scored 95% accuracy at a cost of $8 to run the test. Claude Opus ran the same test at 94.8% accuracy for roughly $2,600. Basically identical accuracy, about 300 times cheaper. If that number holds up, it’s not a marginal improvement. It’s a different price category entirely.
On another test called MRCR v2, which asks the model to find multiple specific pieces of information (“needles”) hidden inside a very long document, SubQ scored 65.9%. GPT-5.4 scored 39%, Gemini 3.1 Pro scored 23%. But Claude Opus 4.6 scored 78% on that test, so SubQ isn’t beating everyone. It’s beating some of the biggest models at a fraction of the cost, but it’s not the top score on every metric.
On SWE-Bench Verified, which tests coding ability, specifically, can the model fix real bugs in real open-source projects, SubQ scored 81.8 against Claude Opus’s 80.8. That’s close enough that you could argue it’s noise, or you could argue it’s a genuinely competitive result from a 13-person company that’s been operating for less than two years.
Why I’m Not Fully Convinced Yet
I spent more time reading the skeptics than the hype, and their concerns are legitimate.
There’s no published research paper. The benchmarks were run under conditions that Subquadratic controlled, sometimes just single runs because of cost. Several AI researchers on X and Hacker News are saying they won’t form an opinion until someone outside the company reproduces the results. Subquadratic says a technical report is “forthcoming.” No date on that.
And the history of this particular claim, we can almost do transformers but cheaper and longer, is uncomfortable. Magic.dev announced a 100-million-token context model back in August 2024 with a claimed 1,000x efficiency advantage. They raised something like $500 million. As of early 2026, there’s no public evidence of that model being used anywhere outside Magic itself. The parallels are hard to ignore. Same context window ambitions, same efficiency framing, same focus on software engineering use cases, same limited external access at launch.
There’s also a more technical concern that the researchers are raising. Past architectures like Mamba and RWKV also promised linear or subquadratic scaling. Both ran into a similar wall: approaches that are theoretically more efficient often underperform standard attention in practice on real benchmarks, especially at the scale where frontier models operate. Or they end up as hybrids anyway, mixing their architecture with regular attention. SSA might be different. But nobody outside the company knows that yet.
Honestly, I don’t know if SubQ clears this bar. That’s not me being cautious that’s just where the evidence sits right now.
What Actually Changes If This Works
So let me explain why people are paying attention even with all the uncertainty, because the practical upside is real.
Right now, if you want to build a product that uses AI to reason over a large body of documents, say, a legal research tool, or a code review system, or anything that needs to read a full knowledge base you basically have to build a RAG pipeline. You need to chunk the documents, embed them (which means converting them into numerical representations a model can search through), store them in a vector database, write a search layer to retrieve relevant chunks, and then stuff those chunks into a prompt before calling the AI. This whole stack exists as a workaround. It’s extra infrastructure, extra cost, extra maintenance, and extra ways for things to go wrong.
If SubQ’s 12 million token context window works the way they claim, not just accepting the tokens but actually reasoning reliably across all of them then for a lot of use cases, you skip the whole stack. Put the documents in directly. Ask the question. Get the answer.
RAG isn’t dead if this works. You still need it for real-time data, frequently updated information, user-specific personalization. But static use casesinternal codebases, legal document review, entire technical manuals, case files those are the obvious early candidates where this changes product design completely. A legal team that can dump an entire case file into one prompt and ask questions across all of it, without worrying about which chunks got retrieved, would pay for that. A software company that can load an entire codebase and ask the AI to find patterns across 80 files at once without setting up a retrieval layer that’s a real product difference.
The price point matters too. If the frontier-level quality is there at 1/300th the cost, that opens up use cases that aren’t economically viable today. Right now, processing a 400-page legal document through Claude Opus might cost more per query than a client is willing to pay for routine work. At SubQ’s claimed price point, you could run it on every document without thinking about it.
The $29 Million Question
I think the architecture is probably real. The SSA mechanism makes conceptual sense. The team has the background to build something like this. What I genuinely don’t know is whether it holds up at scale, whether the reasoning ability matches Opus or GPT-5 on tasks that go beyond long-context retrieval, and whether the benchmark numbers survive independent testing.
The technical report is the moment that matters. If it comes out and those numbers reproduce when researchers outside the company try to check them, this is a big deal maybe the biggest AI launch this year, from a team in Miami with 13 people, not from a lab with thousands of engineers and billions in compute budget. And that itself is worth noting. Some of the most important architecture changes in AI history came from small teams with a specific insight that the bigger players had missed or deprioritized.
But if the report never comes, or the numbers fall apart, SubQ joins Magic.dev and a few others in a growing list of context window announcements that didn’t quite make it to the real world.