Let me tell you about the most interesting release of May 2026, and it didn't come from OpenAI, Anthropic, or Google.
On May 5, a startup called Subquadratic shipped SubQ 1M-Preview — a language model with a native 12 million token context window, built on a fundamentally different architecture than anything currently in production. The company raised $29M in seed funding to do it, and their core claim is simple: this is the first commercially available LLM that isn't a transformer.
That's not a marketing line. That's an architecture claim with significant implications.
Standard transformer attention scales at O(n²) in context length. What that means in practice: double your context window, quadruple your compute cost. That's why long-context models are expensive, why most "1M context" claims come with quiet caveats about quality degradation past a certain point, and why the industry has accepted slow, expensive inference as the price of playing in the frontier.
Subquadratic's claim is that their sparse, subquadratic attention architecture eliminates that quadratic scaling. Their model processes 12 million tokens natively — not with chunking tricks, not with retrieval augmentation, but as a single context window — at roughly 1/5 the cost of comparable frontier models on long-context workloads. They also claim up to 52x faster attention at scale.
Now, a critical note: those numbers are vendor-reported. No independent third party has published SubQ against MRCR, RULER, or the standard long-context benchmarks that would let us verify those claims. I'm treating them as marketing until I see peer-reviewed data. That's the intellectually honest position.
But here's what's worth paying attention to regardless of the exact numbers: this is the first time a commercial entity has shipped a subquadratic-attention model and put real money behind the architecture claim. That alone makes this worth watching.
The LLM industry is in an awkward position. TheIntelligence Index ceiling from April — GPT-5.5 at 60.24 — hasn't been broken in six weeks. The labs that didn't ship in April (Anthropic past Opus 4.7, Google past Flash Lite, Meta past Muse Spark) are presumably building toward their next releases. Nobody is in a hurry to drop a frontier model two weeks after that sprint.
What that means: the action has moved from benchmark chasing to architecture innovation. SubQ is the most visible example, but the broader pattern is that the industry is running into physical limits on transformer scaling and looking for the next path forward. Subquadratic attention, state-space models (Mamba, RWKV), mixture-of-experts optimizations — these are all attempts to break the O(n²) barrier that transformer attention imposes.
SubQ's entry is significant because it's commercial, not academic. We've seen subquadratic architectures in research papers for years. What's new is a company with actual revenue ambitions and a product you can call via API today.
The first release comes with two products: SubQ 1M-Preview (the base model) and SubQ Code, a repo-wide coding agent built to leverage the full 12 million token context. The idea behind SubQ Code is that you can feed it an entire codebase at once and have it reason about architecture, dependencies, and bugs across the full context — without the chunking and retrieval that current code agents require.
The pitch is compelling in theory. Current code agents work by retrieving relevant chunks of code and reasoning over those chunks. The quality of the agent's output is limited by the quality of the retrieval. SubQ's claim is that with 12M tokens of native context, you don't need retrieval at all — you just give the model the whole codebase and let it reason.
Whether that works in practice is the question. Repo-wide code understanding at that scale requires the model to maintain coherent understanding across millions of tokens of context, which is a different capability than the short-context reasoning that most benchmarks test. Until we see production evidence, I'm skeptical but curious.
I want to be direct about what we don't know yet.
First: the performance claims. 52x faster attention and 1/5 the cost are vendor figures. Independent benchmarks don't exist yet. The history of AI performance claims includes plenty of selective reporting, and I'd want to see peer-reviewed verification before I updated my mental model of what subquadratic attention can actually deliver.
Second: the architecture itself isn't novel. Subquadratic attention as a research area has been active for years — Mamba, RWKV, Hyena, and others have all explored non-transformer architectures. What's new is commercial deployment at scale. The gap between research demonstration and production reliability is significant.
Third: the model is Preview. That usually means limited availability, potential instability, and features that will change. SubQ is not a finished product — it's the first version of something that may become significant.
Here's the thing I'm actually watching: if subquadratic attention works at commercial scale, it breaks the economic model that the entire LLM industry has been built on.
The current model is: better models cost more to run, more context costs more, and the primary competitive dimension is capability. If SubQ or a competitor achieves the same quality at 1/5 the cost, that changes every pricing conversation, every infrastructure decision, and every architectural choice in production AI systems.
We're not there yet. But May 2026 just got interesting.
*SubQ 1M-Preview: 12M token context, subquadratic attention, $29M seed, API available now. Vendor claims require independent verification, but the architecture direction is worth watching.*