Miami Startup Unveils AI Model With 12M-Token Context, Outperforming Google and Anthropic

A Miami-based startup called Subquadratic has launched its first AI model with a 12-million-token context window. The company claims its technology outperforms leading models from Anthropic and Google on key benchmarks while running far more efficiently. Subquadratic’s model uses a new architecture called Subquadratic Selective Attention (SSA). Unlike traditional dense attention, this approach scales linearly in compute and memory as context length grows. The result is a system that processes information 52 times faster than standard models at a million tokens.

The startup’s model has already set new records. It scored 83% on the multi-reference retrieval benchmark MRCR v2, surpassing GPT-5.5’s 74%. On SWE-bench, a coding-focused test, it achieved 82.4%, outperforming both Anthropic’s Opus 4.6 and Google’s Gemini 3.1 Pro. Subquadratic is rolling out its technology through three products: an API with a 12-million-token window, SubQ Code (a command-line agent for developers), and SubQ Search. The company has also announced plans for a future model with a 50-million-token context window. Backed by $29 million in funding, Subquadratic is now valued at $500 million. Its launch comes at a time when many frontier AI models in 2026 advertise million-token context windows—but few use them effectively.

Subquadratic’s model delivers both speed and accuracy, setting new benchmarks in retrieval and coding tasks. With a $500 million valuation and plans for even larger context windows, the startup is positioning itself as a major player in efficient AI processing. The company’s API and developer tools are now available in beta for early users.