SubQ Snap: Long-Context Decision Model for Enterprise Agents
On October 12, 2026, Subquadratic launched SubQ Snap, a multimodal decision model built specifically for high-frequency, bounded decisions inside enterprise agentic workflows. The official BusinessWire release describes it as “the first long-context decision model purpose-built for enterprise.” The launch marks a concrete step toward separating the two classes of work inside an AI agent loop: generative reasoning that requires a large frontier model, and structured decision-making that does not.
SubQ Snap is available today in public preview at $0.05 per million input tokens, with output tokens provided free.
The Problem SubQ Snap Is Solving
Every production AI agent makes dozens of structured decisions before it completes a single task. Which file in a large codebase is relevant? Which tool should be called next? Does this test result pass? Should the current step be escalated to a human? The answer to each question comes from a small, predefined set of options. But most enterprise agents route these calls to the same frontier LLMs handling open-ended generation, code synthesis, and complex reasoning.
That is expensive by design. A frontier model call can cost 100 to 1,000 times more per request than a specialized decision model, and the latency runs from several seconds to several minutes. In a workflow that chains 30 to 50 such decisions, the arithmetic compounds quickly. The cost of running agents at enterprise scale is, in large part, the cost of using overqualified models for underqualified decisions.
Decision models are a direct response to this inefficiency. They accept a state description, a question, and a fixed set of allowed answers, then return probabilities over that answer space without generating any text. The cost-per-decision falls by orders of magnitude. The challenge has been context: most existing decision models top out at 32K tokens, which means they cannot see the full artifact when the relevant evidence is distributed across a long codebase, a contract with many exhibits, or a multi-session interaction history.
SubQ Snap is built specifically to remove that constraint.
What SubQ Snap Does
SubQ Snap combines two architectural pieces that address the two sides of decision-model workload.
The first is Subquadratic Sparse Attention (SSA), the company’s core research contribution. Standard transformer attention computes a relationship between every pair of tokens, which means the compute cost grows quadratically as context length increases. Doubling the context quadruples attention compute. SSA replaces dense attention with a content-dependent sparse selection that grows linearly. At 12 million tokens in research testing, SSA reduced attention compute by nearly 1,000x compared with standard transformer architectures.
The second is a decision head that maps the model’s contextual representations directly to a probability distribution over the predefined answer options, bypassing autoregressive text generation entirely. The result is a model that can read a 384K-token artifact and return a structured decision in about one second, including 100K tokens of input.
Key specifications:
- Maximum context: 384K tokens of mixed text and images
- Throughput: 100K-token request processed in approximately 1 second
- Modality: text and images, enabling reasoning over screenshots, diagrams, charts, and scanned documents alongside written text
- Pricing: $0.05 per million input tokens; output tokens are free
- Availability: public preview, limited capacity, with a waitlist as demand scales
Benchmark Performance
Subquadratic evaluated SubQ Snap across the Decision Index (v0.2.1), the public benchmark for structured decision models, and against a subset of 16 tasks mapped to common enterprise workflows.
| Benchmark | SubQ Snap | Jev 1.13 |
|---|---|---|
| Decision Index 0.2.1 (full index) | 61.7 | 57.9 |
| Decision RULER at 1M tokens | 96.3 | N/A |
| Decision LCR (real multi-doc, 72K to 115K tokens) | 74.0 | 48.0 (chunked) |
| Enterprise workflow average (16 benchmarks) | 67.6 | 61.6 |
| Phishing detection (PhishNChips) | 47.1 | 25.1 |
| Contract NLI (full-NDA clause review) | 70.7 | 59.1 |
| Financial entity sentiment (FinEntity) | 92.8 | 80.8 |
| Code execution prediction (CRUXEval) | 73.8 | 57.1 |
| Hallucination detection (RAGTruth) | 64.8 | 51.3 |
Across the 16 enterprise-specific benchmarks, Snap leads on 13 of 16 tasks. The Decision LCR comparison is especially significant: Jev required chunking at 72K to 115K tokens and still scored 26 points lower than Snap processing the same documents in a single call.
Four Roles in a Running Agent Workflow
Subquadratic identifies four recurring patterns in enterprise agent workflows where decision models can take work away from frontier LLMs.
Routing. Which team, tool, or model should handle this request? Routing decisions require reading enough context to classify intent accurately, but the answer space is small and fixed. A decision model is well-suited to handle this at millisecond latency.
Tool picking. Which of the available tools should the agent call next? At long context, the relevant evidence can span hundreds of prior tool calls and the content they returned. Snap processes that full history without truncation.
Checking. Does this result meet the acceptance condition? Did the test pass? Is this output within the specified format? These are binary or small-category decisions that occur at every step of a workflow.
Stopping. Should the agent continue, ask for clarification, or escalate to a human? Stopping decisions carry the highest stakes in production systems. A model that can see the full context of a multi-session interaction is more likely to reach the correct stop condition.
Across all four patterns, the key variable is whether the decision model can actually see the evidence it needs. With a 384K-token window, Snap extends that reach substantially.
Enterprise Implications
For teams operating AI agents at scale, the cost calculus is shifting. A workflow that calls a frontier LLM 40 times per task at an average of $1 per million tokens costs roughly 40 times more per decision than a workflow that routes most of those calls to a decision model at $0.05 per million tokens. The frontier model becomes a specialist reserved for code generation, complex reasoning, and synthesis.
This is not a hypothetical architecture. Several enterprise agent builders have already published analyses showing that 60 to 80 percent of the LLM calls in a typical agentic workflow are bounded decisions that do not require open-ended text generation. If that share can be rerouted to a capable, fast, low-cost decision model, the economics of running agents at enterprise scale change materially.
SubQ Snap also addresses the latency side of the equation. A 100K-token decision in approximately one second, compared with several seconds to several minutes for a frontier model call on the same context, changes what is practical in synchronous human-facing workflows. Customer support routing, real-time document review, and interactive code analysis all become viable at production latency.
The multimodal capability adds a further dimension. Enterprise workflows increasingly operate over mixed artifacts: a contract alongside its annotated PDF, a codebase alongside CI screenshots, a financial report alongside chart images. A decision model that accepts text and images in the same 384K-token call can reason over that full artifact without preprocessing steps that introduce errors and latency.
The Decision Model Ecosystem Is Taking Shape
SubQ Snap enters a category that now includes multiple players. TypeSafe AI’s Jev model launched in September 2026 and raised $870 million at a $7.5 billion valuation. Perplexity’s Decider and Laya are active in specific routing markets. Jurin AI launched JJ-1 on October 13, targeting Japanese enterprise workflows.
What distinguishes the current moment is that decision models are moving from research novelties to production infrastructure. The arrival of multiple funded, competing products in 2026 follows the same pattern that defined the LLM-as-a-service market two years earlier: a new capability class that starts as a single notable product, then attracts capital and competition as the enterprise use case becomes clear.
The limiting factor for previous decision models was always context. Routing a customer support ticket is tractable at 32K tokens. Routing a complex multi-step software engineering task across a 200K-token codebase with accumulated tool call history is not. SubQ Snap is making a specific architectural bet that resolving this context constraint is where the market breaks open.
For enterprise AI teams building production agent systems, the near-term decision is straightforward: audit the LLM calls inside your existing agent loops and identify what fraction are bounded decisions. If that fraction is high, a hybrid architecture pairing a frontier generative model with a capable, long-context decision model is worth evaluating now. SubQ Snap provides one of the more technically differentiated options in that category as of October 2026.
Subquadratic has also announced that SubQ 1.1 Small, a full language model built on the same SSA architecture, is in early access with selected design partners, signaling that the subquadratic approach is being extended beyond the decision layer toward the full generative stack.
For teams evaluating the economics of enterprise agentic workflows and looking to control the AI token costs that scale with agent volume, SubQ Snap represents a technically substantive option that arrived this week. Reach out to Enera to discuss how decision model routing fits into your agent architecture.