Subquadratic's SubQ Snap 1.1 Ties for Top AI Decision Model With 384K Context

Artificial Analysis now ranks Subquadratic’s SubQ Snap 1.1 alongside d1-ultra and Elephant 9B at the top of its Decision Index v1.0. Each model scores 65, while the next-ranked model, Jev 1.14, scores 57.

The shared score conceals substantial differences in context capacity, cost, accuracy, and willingness to abstain. Those differences determine which model fits a given production workload.

One score, three risk profiles

Artificial Analysis calculates the index from four equally weighted evaluations covering knowledge and hallucination, reasoning, tool calling, and classification. Decision models select from a fixed set of options and return a probability for each in one call. Their probability outputs also let teams measure calibration, meaning how closely confidence tracks observed accuracy.

On AA-Omniscience-Decision, the index’s knowledge and hallucination component, the three leaders follow distinct strategies:

• Elephant 9B attempts 18% of questions, records 12% accuracy, and posts a 93% non-hallucination rate.
• SubQ Snap 1.1 attempts 32%, records 20% accuracy, and posts an 86% non-hallucination rate.
• d1-ultra attempts 42%, records 32% accuracy, and posts an 85% non-hallucination rate.

The non-hallucination rate measures how often a model abstains among questions it does not answer correctly. Elephant 9B supplies the fewest answers and avoids the most incorrect ones. d1-ultra attempts the most questions and achieves the highest accuracy, while SubQ Snap 1.1 falls between them on both measures.

Snap stretches the context window

Model specifications create the clearest separation. SubQ Snap 1.1 supports a 384K-token context window, compared with 32K for d1-ultra and 16K for Elephant 9B. Snap therefore offers 12 times d1-ultra’s capacity and 24 times Elephant 9B’s.

Subquadratic attributes that capacity to Subquadratic Sparse Attention, or SSA. Standard dense attention requires computation that grows quadratically with input length. The company says SSA makes its attention computation grow linearly, reducing the cost of processing long inputs.

Snap also uses a decision head that maps the model’s internal representation directly to probabilities over the permitted answers. This avoids generating a response token by token. Subquadratic reports that the model can process 100K tokens and return a decision in roughly one second.

Long context retains its benchmark score

On HLE-MultiChoice, the text-only multiple-choice subset of Humanity’s Last Exam, SubQ Snap 1.1 and Elephant 9B each score 18%. d1-ultra scores 15%. The result gives Snap a three-point lead over d1-ultra on this reasoning evaluation, though a single benchmark cannot establish reasoning quality across every workload.

Short inputs favor Elephant’s price

Elephant 9B costs 72% less per million tasks than SubQ Snap 1.1 and about 74% less than d1-ultra under the published estimates. It offers the lowest listed cost for workloads that fit within 16K tokens and can tolerate its high abstention rate.

Subquadratic separately lists Snap’s input price at $0.04 per million tokens. At that rate, a 100K-token input carries an input charge of $0.004 before any other fees. Actual spending therefore depends on input length, and the table’s per-task estimate should not be treated as the price of every long-context request.

Inputs longer than 32K tokens leave Snap as the only tied model that can accept the full prompt in one call. Deployments using d1-ultra or Elephant 9B would need to retrieve, summarize, or split material before making a decision, adding engineering work and potentially excluding relevant context.

Where decision models fit in an agent stack

Agent systems repeatedly make bounded choices that do not require a generated explanation. Common examples include:

• selecting relevant files from a repository;
• choosing the next tool or workflow step;
• classifying a test result as passing or failing;
• routing uncertain cases to a larger model or a human reviewer.

Each decision adds latency and cost to an agent loop. A specialized model can handle these structured calls directly, while probability outputs allow developers to set confidence thresholds and escalate uncertain cases. Snap’s larger context window extends that approach to decisions involving long codebases, document collections, or interaction histories.

A benchmark tie, three deployment choices

The aggregate tie leaves the deployment choice to workload constraints. Elephant 9B offers the lowest published cost and the most selective answer policy within 16K tokens. d1-ultra records the highest knowledge accuracy and attempt rate within a 32K window. SubQ Snap 1.1 provides a middle-ground abstention profile and a 384K context window for decisions that require substantially more source material.