Subquadratic launches SubQ Snap, a long-context decision model it says outperforms Jev

Subquadratic, a startup focused on optimizing how foundation models scale, announced on October 12, 2026 the release of SubQ Snap, a specialized decision model built for agentic AI workflows. Unlike traditional large language models designed for open-ended text generation, Snap is engineered to handle the data-intensive but structured decisions that arise constantly in automated systems: determining which files matter, selecting the right tool for a task, evaluating test results, or routing queries to more capable frontier models. The company positions this as a cost and latency reduction strategy for enterprises running complex agent systems.

The technology rests on Subquadratic's earlier innovation, Subquadratic Sparse Attention (SSA), which transforms attention compute cost from quadratic to linear complexity as context length grows. This architectural breakthrough enables Snap to process longer inputs without the exponential compute penalty typical of standard transformers. Combined with a decision head that maps representations directly to probability distributions over allowed answers, bypassing slow autoregressive generation, the model avoids the usual trade-off between context size and response speed. In benchmark tests across 16 Decision Index categories spanning customer support, security, legal review, finance, and code tasks, SubQ Snap averaged a score of 67.6 compared to competitor Jev's 61.6, leading on 13 of 16 specific tasks including phishing detection (+22.0 points) and contract clause review (+11.6 points).

Pricing reflects Subquadratic's efficiency claims at just $0.05 per million input tokens with output tokens free, a significant departure from traditional LLM billing models. While Snap is not positioned as a general-purpose chatbot but rather as infrastructure for enterprise agent workflows, the benchmark results suggest companies deploying sophisticated AI automation may find meaningful cost reductions without sacrificing decision quality. The launch comes as enterprises increasingly seek specialized models optimized for specific workflow stages rather than relying on monolithic foundation models for all tasks.