Subquadratic launches SubQ Snap: a 384K-context decision model for agent routing, tool picks and checks at $0.05/M input

Why it matters

Agents make many bounded decisions every turn: which files are relevant, which tool to call next, whether a test passed, and whether a step needs a frontier model. Sending all of these to a generative LLM is slow and costly. Existing decision models (TypeSafe's Jev, Perplexity's Decider) mostly stop around 32K tokens, so teams fall back on retrieval, truncation or chunking and lose evidence such as a constraint stated 40 turns earlier. Subquadratic's SubQ Snap, released on October 12, 2026, targets that gap.

How it works

Snap combines Subquadratic Sparse Attention (SSA), which selects attention by content so compute grows linearly with context, with a decision head that maps representations directly to a probability distribution over the allowed answers instead of decoding text. It accepts up to 384K tokens of mixed text and images, so screenshots, charts and scanned pages can sit next to logs and code.

Numbers (vendor-reported)

• Decision Index 0.2.1: 61.7 vs Jev 57.9; leads in four of five areas.
• 16 enterprise benchmarks: 67.6 vs 61.6 average, ahead on 13. Big gains on PhishNChips (+22.0), CRUXEval (+16.7), RAGTruth (+13.5); behind on When2Call (-7.6), API-Bank (-1.6) and ACOS (-0.6).
• Decision LCR (72K–115K multi-document): 74.0 one-shot vs Jev 48.0 with chunking.
• Decision RULER: 96.3–99.8 from 32K to 1M tokens.
• Latency per the launch post: ~80 ms minimal context, 250 ms at 32K, 660 ms at 100K, 2.5 s at 384K.
• Price: $0.05 per million input tokens, output free.

Getting started

Create a key at console.subq.ai and call POST https://api.subq.ai/v1/systemone with model: subq-snap-preview, a state, and typed questions (noul, choice, score). The Decisions API is not OpenAI-compatible. Access is a first-come public preview with a waitlist once capacity runs out. A practical pattern: route tool selection, result checks and stop decisions in a coding agent to Snap, keep code synthesis on a generative model, and validate the vendor scores on your own labeled traces before relying on it in a latency-critical path.