- Subquadratic exited stealth May 5 with $29M raised at a $500M valuation, led by Javier Villamizar and Tinder co-founder Justin Mateen.
- SubQ ships a native 12-million-token context window, roughly 12x larger than every commercial frontier model on the market today.
- A new architecture called Subquadratic Selective Attention (SSA) runs 52x faster than FlashAttention at the 1M-token mark and scales linearly, not quadratically.
- On a long-context workload that costs roughly $2,600 on Opus 4.6, SubQ finishes for $8 — and scores higher on RULER 128K and MRCR v2.
Why is every long-context benchmark suddenly being rewritten by a nine-figure startup that almost nobody had heard of two weeks ago? On May 5, 2026, a stealth lab called Subquadratic walked out of the shadows with $29 million in funding, a $500 million valuation, and a model that does something Transformer-based frontier labs have spent four years insisting was impossible at scale: a 12-million-token context window that scales linearly in both compute and memory. The pricing math is the part that has founders staring at their monthly OpenAI bill. One representative long-context workload that costs roughly $2,600 on Claude Opus 4.6 runs to completion on SubQ for about $8. That is not a typo. That is a 325x gap on a single task. And the benchmarks say SubQ is the more accurate of the two.
The Architecture That Broke The Quadratic Wall
What Subquadratic Selective Attention Actually Does
Standard Transformer attention is quadratic in sequence length. Double the context, quadruple the compute and memory. That is why every frontier lab has been chasing tricks — sliding windows, mixture of experts, KV-cache compression — to fake longer context windows without actually paying the quadratic tax. Subquadratic’s bet, codified in a proprietary architecture they call Subquadratic Selective Attention or SSA, is that the right answer is to throw out the quadratic operation entirely.
SSA is a fully sparse attention system. Compute grows linearly with sequence length. Memory grows linearly with sequence length. The published number that matters is this: at the 1-million-token mark, SSA runs 52 times faster than FlashAttention, which is the most heavily optimized quadratic-attention kernel in production. The 12-million-token window is not a marketing slot you can hit by stitching documents together. It is native context the model attends to in one pass.
Trend Insight — Every frontier lab is now staring at a forced choice. Either acquire a sub-quadratic stack, license one, or watch a nine-person startup undercut frontier pricing on the workloads enterprise customers care about most: codebase indexing, multi-document research, and long-horizon agent runs.
The Benchmarks Frontier Labs Did Not Want To See
RULER, MRCR, And A 325x Cost Gap
A long context is only worth something if the model actually attends to it. The two evals that have become industry shorthand for that question are RULER and MRCR. On RULER at 128,000 tokens, SubQ scored 97 percent. Claude Opus 4.6, which until now has been the unspoken benchmark for serious long-context work, scored 94 percent on the same test. Opus had to load the full 128K context to do it. SubQ barely sees that as a warmup.
The multi-needle retrieval benchmark MRCR v2 is where the gap turns embarrassing for the incumbents. SubQ scored 83. Opus 4.6 scored 78. GPT-5.4 scored 39. Gemini 3.1 Pro scored 23. On a workload that mixes retrieval and reasoning across millions of tokens, two of the three biggest names in AI are not even in the same league. Pair the accuracy numbers with the pricing — $8 versus roughly $2,600 for the same end-to-end task — and the conversation about long-context AI looks completely different than it did a month ago.
Trend Insight — Quadratic attention has been the silent tax that propped up frontier-lab margins on every long-context API call. Once a sub-quadratic competitor proves it ships and benchmarks higher, that margin starts evaporating, and the legal department at every closed lab is about to be very busy.
What This Unlocks For Developers
The API, The CLI Agent, And The 50M-Token Target
Subquadratic is shipping two products in beta. The first is a hosted API that exposes the full 12-million-token window — no chunking, no retrieval augmentation, no janky middleware. You hand the model a context, and it reads the whole thing. The second is SubQ Code, a CLI agent built on the same model, aimed directly at the autonomous coding niche that Anthropic’s Claude Code and OpenAI’s Codex CLI have been splitting between them.
The roadmap is louder than the launch. The team has publicly committed to a 50-million-token context window in Q4 of this year. To put that in perspective, 50 million tokens is roughly the entire source code of a large enterprise monorepo plus its test suite plus its design docs plus a year of internal pull-request discussion — all loaded into one inference call, with no retrieval system in front of it. If they hit that target, the entire RAG-vs-long-context debate that has dominated 2025 quietly becomes a footnote.
Trend Insight — The most underpriced engineering bet of the next twelve months is not building another agentic wrapper around GPT-class APIs. It is rewriting your data pipeline so it can hand a model the whole context — because that capability just got 325x cheaper, and the lab that did it is hiring.
Related
- SAP Just Bet the Whole Company on 200 AI Agents
- Wall Street Just Crowned Nvidia’s Newest Threat
- Alphabet Just Quietly Bet $2.1B on Curing Every Disease
- $650M for AI That Rewrites Its Own Code
Sources
- SiliconANGLE — Subquadratic launches with $29M to bring 12M-token context windows to AI
- The New Stack — Subquadratic debuts a 12-million-token context window
- eWeek — Subquadratic Launches SubQ, a 12M-Token AI Model for Long-Context Tasks
AI Biz Insider · AI Trends EN · aibizinsider.com
답글 남기기