Everyone in AI is arguing about benchmark scores. Subquadratic is arguing about architecture. That's the more interesting fight.
SubQ 1M-Preview dropped May 5, 2026, with $29M in seed funding and one claim that should make every infrastructure engineer pay attention: this is the first commercially available LLM where compute scales linearly with context length, not quadratically. They call it subquadratic attention. It's not a transformer under the hood.
Let's talk about why that matters, what the benchmarks actually show, and why you should care even if the vendor numbers are inflated.
Standard transformer attention has a fundamental cost structure: every token attends to every other token. Double your context length, and you quadruple the compute required. Not roughly double — actually quadruple. That's the quadratic tax, and it's why "1M token context" sounds impressive until you look at what it costs to actually use it.
The industry response to the quadratic tax has been a decade of workarounds. RAG pipelines that retrieve a small number of relevant chunks and stuff them into a manageable context window. Chunking strategies, semantic splitting, hierarchical retrieval. Engineers spend more time building retrieval infrastructure than building the actual product. And all of it exists because transformers get expensive fast as context grows.
Subquadratic's bet is that the tax is optional. If you can design attention that scales linearly — where compute grows at the same rate as context — then the entire retrieval layer becomes unnecessary for tasks that genuinely need millions of tokens of context.
SubQ 1M-Preview uses sparse, subquadratic attention end-to-end. This isn't a modification to transformer attention — it's a different architectural approach. The company has published benchmarks on RULER 128K (a standard benchmark for reasoning over extended inputs) showing 95.6% accuracy versus 94.8% for Claude Opus 4.6. They've also published a third-party-verified MRCR v2 score of 65.9, comparing favorably with GPT-5.5 at 74 and Claude Opus 4.7 at 32.2.
On SWE-Bench Verified — a benchmark for software engineering tasks that tests a model's ability to actually work with code — SubQ scored 81.8 versus Opus 4.6 at 80.8 and DeepSeek 4.0 Pro at 80.0.
These are real numbers. They're also vendor numbers. The caveat that applies to every unverified benchmark applies here: until independent third parties run these tests, treat them as directional, not definitive.
What is independently verifiable: SubQ's architecture reduces attention compute by roughly 1,000x compared to standard transformer approaches at 12M tokens. The vendor claims 52x faster attention versus FlashAttention in their architecture-level comparison, with 63% less compute required. Those claims are consistent with the academic literature on subquadratic attention methods — Mamba, RWKV, Hyena have all shown that linear or subquadratic scaling is achievable. The question has always been whether you can maintain accuracy while doing it.
Native 12M token context. Not interpolated. Not extended with degraded quality past some threshold. The architecture is designed for it from the ground up.
For context: most models that advertise 1M token context windows start showing significant quality degradation well before they hit that ceiling. The attention mechanism simply can't maintain coherent reasoning over that much context at transformer quadratic cost. SubQ's architecture sidesteps this by not having the quadratic cost in the first place.
The product lineup reflects this: SubQ Code is a coding agent built to load entire codebases into a single context window and work across the full repository in one pass. SubQ Search is a long-context search tool with Deep Research capabilities at chatbot speed. The API is for developers who want to build on top of the full context directly.
The vendor claims roughly 1/5 the cost of frontier models on long-context workloads. That's the number that matters for production systems.
If you're running RAG pipelines today, you know the cost structure: embedding generation, retrieval search, context assembly, model inference on the retrieved chunks. For tasks that require reasoning over large document corpora — legal discovery, codebases, research archives — the retrieval overhead is significant and the context assembly quality is variable.
SubQ's value proposition: skip the retrieval layer entirely. Put the entire corpus in context. Pay linear compute instead of quadratic compute. Get results faster and cheaper.
Whether that claim holds in production is the right question. The architecture is real. The benchmarks are published. The $29M in funding suggests investors found the claims credible enough to back. The next six months will tell us whether SubQ can generalize those numbers outside the carefully constructed benchmark environment.
Subquadratic attention as a research area is not new. Mamba (2023), RWKV (2023), Hyena (2023) — all demonstrated that subquadratic scaling was achievable. The consistent failure mode across all of them: accuracy at the tasks that matter. Linear scaling is easy if you don't care about quality. Getting both has been the hard part.
SubQ's RULER and SWE-Bench numbers suggest they've gotten past that failure mode. That's significant if it holds. But the benchmark suite for long-context tasks is thin, and the tasks that matter most in production — real-world document reasoning, multi-document synthesis, complex codebases with ambiguous requirements — don't map cleanly to synthetic benchmarks.
The other honest concern: $29M seed is early stage. The gap between "works in the paper" and "reliable production service" is where most AI infrastructure bets die. SubQ has to execute on serving infrastructure, API reliability, and the support cadences that enterprise buyers demand. That's a different skill set than model architecture.
If SubQ's claims hold up under independent scrutiny, the implication is significant: the transformer quadratic tax is breakable, and the entire RAG-heavy AI engineering stack that exists to work around it is a transitional artifact, not a permanent structure.
That's a big if. But it's the right kind of if — one that, if answered correctly, changes how we build AI systems fundamentally.
The practical bet for teams working with large context today: watch SubQ's third-party benchmark results closely. If RULER, MRCR, and SWE-Bench numbers hold across independent evaluation, the architecture is worth building around. If they don't, it's another promising paper that didn't translate.
Either way, the question SubQ is asking — what if we just fixed the scaling problem instead of working around it? — is the most interesting question in AI infrastructure right now.
*SubQ 1M-Preview released May 5, 2026 by Subquadratic. $29M seed. 12M token native context. API and SubQ Code (CLI) available via private beta. Benchmarks: RULER 128K 95.6%, SWE-Bench Verified 81.8%, MRCR v2 65.9 (third-party verified). Vendor claims ~1/5 cost of frontier on long-context workloads.*