Subquadratic released SubQ Snap 1.1 on October 22, 2026, updating its hosted decision model for long-context agent workflows. The announcement emphasizes better uncertainty calibration and action selection, with native 384K-token context and a price of $0.04 per million input tokens. It is a version update to Snap, whose original announcement was October 12.
What a decision model changes
The original Snap explanation describes a constrained answer space: an application provides context, a question and allowed answers, and receives scores for those answers. That differs from asking a general language model to write an explanation and then extracting a decision from its prose.
Subquadratic combines a decision head with its Sparse Attention architecture, which is designed to make reading long context more efficient. The rationale is relevant to agents: choosing one of three actions can still depend on information scattered across a repository or conversation. A short answer does not mean the evidence needed to choose it is short.
The architectural distinction matters, but it should not be interpreted as permission to attach every available document to every request. Our assessment is that context selection still has a business purpose: exclude unrelated records, preserve resource ownership and include the evidence that actually governs the decision.
The release focuses on knowing when not to act
Subquadratic reports that Snap 1.1 improves When2Call accuracy from 75.2 to 96.2 in its action-arbitration evaluation. It also reports reducing expected calibration error on held-out probability problems from 0.108 to 0.014. Those are vendor-reported evaluation results, not guarantees for a customer’s data.
The release is unusually useful about a weakness: it says Jev 1.14 retains a substantial advantage on its phishing-classification comparison, attributing much of the gap to threshold selection. Similar discrimination quality can therefore produce different operational results when a threshold turns probabilities into actions.
A probability becomes useful only within a policy
Consider a support agent deciding whether to call a refund tool, ask for missing details or hand a case to a person. A high score for “refund” is not proof that the caller owns the order or that a refund is permitted. The application must verify those facts through its authoritative systems.
A meaningful evaluation would separate three failures: acting when information is missing, asking unnecessary questions when the request is complete, and choosing the wrong action despite complete evidence. The costs differ. An extra clarification may be tolerable; an unauthorized payment is not. One aggregate accuracy score hides that distinction.
Thresholds should reflect those consequences. Keep a set of cases near the decision boundary and include unfamiliar requests where abstention is desirable. Examine whether scores remain useful as case mix changes, rather than treating a confidence value as a permanent measure of certainty.
Long context should earn its cost
Snap’s original launch argued that an agent may lose decisive details when a long artifact is reduced into fragments. That is a reasonable hypothesis to test on a concrete workflow. For a contract triage task, compare decisions using the full permitted document set with decisions using the team’s current retrieval method. Review where an exception clause changes the answer.
Measure review time and missed exceptions alongside request cost. The October 22 price is lower than the original Snap rate, but repeatedly reading a large history can still dominate the workload. A cheaper decision that causes more rework may not improve the process.
Snap 1.1 makes a focused case for specialized models inside agent systems. The practical opportunity is to route bounded choices through a scored interface while keeping authority, evidence and escalation explicit. Its calibration claims justify closer evaluation, and its disclosed threshold weakness explains why that evaluation must use the customer’s actual decisions.