Subquadratic's 12M-Token LLM Breaks Context Limits

A Miami-based startup, Subquadratic, has launched an AI model with a 12-million-token context window. This model performs 52x faster than FlashAttention at 1M tokens and costs roughly 1/5 as much as current frontier models, according to Thenewstack and Theneurondaily. For a decade, large language models have been held back by a mathematical bottleneck limiting context windows and inflating costs. Subquadratic claims to have solved this, delivering superior performance at a fraction of the price, as reported by MIT Technology Review. The breakthrough positions companies to leverage AI for complex, long-form tasks previously deemed impractical or too expensive, potentially disrupting established AI giants and accelerating innovation across industries.

Unpacking SubQ's Performance Advantages

Subquadratic's model demonstrates notable performance across key benchmarks. It achieves 92.1% on needle-in-a-haystack retrieval at 12 million tokens and scores 83 on MRCR v2, surpassing OpenAI by nine points, according to Thenewstack. Furthermore, it scores 82.4% on SWE-bench, outperforming Anthropic’s Opus 4.6 (81.42%) and Google’s Gemini 3.1 Pro (80.6%). Critically, SubQ offers this 12M-token context window at roughly 1/5 the cost of frontier models, as detailed by Theneurondaily. The figures suggest a significant challenge to the current market dominance of expensive, context-limited frontier models. Companies relying on existing LLM infrastructure must now re-evaluate their strategies, or risk falling behind.

How SubQ Solved the LLM Scaling Challenge

Subquadratic’s solution, SubQ, is a 12M-token LLM built on a fully sub-quadratic architecture, according to Theneurondaily. The Subquadratic architecture (SSA) scales linearly with input length, directly addressing the scaling limitations that have constrained previous models. Furthermore, the SSA runs 52x faster than FlashAttention at 1M tokens. The combination of linear scaling and speed enables efficient processing of extensive datasets, demonstrating that algorithmic innovation can indeed disrupt the compute-heavy dominance of larger tech companies.

A 12-million-token context window fundamentally alters the trade-offs between context length, performance, and cost in LLM development. The expanded capacity, combined with 92.1% retrieval accuracy, unlocks new possibilities for complex, long-form AI applications previously deemed cost-prohibitive or technically unfeasible. Fields like legal tech, software development, and scientific research can now leverage AI for tasks involving massive datasets, such as analyzing entire codebases or extensive legal documents within a single model. The critical question remains: how quickly will industries adapt to this newfound capability, and what unforeseen challenges might emerge from such extensive contextual processing?