SubQ Launches: 12-Million-Token Context Window That Beats GPT-5.5

What Is SubQ and Why Is It Different?

Subquadratic, a Miami-based startup, launched in May 2026 with SubQ — an AI model boasting a 12-million-token context window. That's roughly 9 million words or about 120 books in a single prompt. The secret is their proprietary Subquadratic Selective Attention (SSA) architecture.

The Problem With Current Context Windows

Traditional transformers use dense attention: every token gets compared to every other token. With 1,000 tokens, that's 1,000² = 1,000,000 comparisons. Scale to millions of tokens and compute costs become astronomical. This quadratic scaling problem has constrained every transformer since 2017.

How SSA Solves the Scaling Problem

SubQ uses content-dependent selection to identify which token relationships actually matter — and only processes those. Unlike DeepSeek's Native Sparse Attention, where the indexer was still quadratic, SSA's selection mechanism scales linearly in both compute and memory.

The Performance Numbers

The benchmarks are striking: 52× faster than dense attention (FlashAttention) at 1 million tokens, and 92.1% accuracy on needle-in-a-haystack retrieval at 12M tokens — a length no other frontier model even approaches.

FAQ

Q: What can I do with 12 million tokens? A: You could load entire codebases, full research paper collections, or complete legal document sets into a single prompt for analysis.

Q: Is SubQ available now? A: The company just launched with seed funding. Availability and pricing haven't been announced yet.

Q: How does this compare to Gemini's context window? A: Gemini's context window is measured in millions of tokens, but SubQ claims both longer context and better retrieval accuracy at extreme lengths.

Stay ahead of the AI curve. Follow @AiForSuccess for daily insights.