Subquadratic, a Miami‑based startup, unveiled a model that shatters the 1‑million‑token ceiling, offering a 12‑million‑token context window. Its Subquadratic Selective Attention (SSA) architecture claims linear scaling in compute and memory, promising 52× speed over dense attention at one million tokens. The launch follows a trend of models stuck at one‑million limits.

Benchmarks on MRCR v2 show SSA scoring 83, nine points ahead of OpenAI’s GPT‑5.5. On needle‑in‑a‑haystack at 12 million tokens, Subquadratic hits 92.1 %. The company also tops the SWE‑Bench verified score at 82.4 %, edging Anthropic’s Opus 4.6 and Google’s Gemini 3.1 Pro, while remaining under a third of the compute cost of comparable models today.

Subquadratic’s 11 Ph.D.‑led team claims the SSA design sidesteps the quadratic bottleneck by selecting content‑dependent positions without a costly indexer. This pure subquadratic mechanism delivers 7.2× speed at 128K tokens and 52.2× at 1M, according to the single‑run benchmarks highlighted in the technical paper to demonstrate scalability across diverse workloads and.

The startup, now valued at $500 million after a $29 million round, offers the 12‑million‑token API and a coding agent called SubQ Code via neoclouds, bypassing major hyperscalers. With a planned 50‑million‑token window for Q4, Subquadratic positions itself as a serious challenger to mainstream LLMs without compromising performance and efficiency in real‑world deployments today.