# SubQ Snap 1.1 Comes to Baseten: Deployment, Context and Pricing Details

The long-context decision model gets another deployment route, but its direct API price should not be confused with dedicated hosting costs.

By Paras Madan | 2026-11-03


Baseten announced support for Subquadratic's SubQ Snap 1.1 on November 2, adding a dedicated-deployment route for the long-context decision model. The model already existed before this announcement. The new development is where teams can deploy it, rather than a new version or a newly announced context window.

## What is available on Baseten

Baseten's [model listing](https://www.baseten.co/library/subq-snap-11/) offers dedicated hardware through a contact-led deployment flow. It describes a proprietary model for routing, ranking, classification and other structured judgments, supporting up to 384K tokens of context and up to four images.

This is different from a chatbot that writes a long answer. Snap returns probabilities, category choices and scores that application code can use to decide what happens next.

The [November 2 announcement](https://x.com/baseten/status/2117356128631976142) confirms the provider rollout. The library page does not publish a dedicated-hosting price or establish that a self-serve, per-token Baseten endpoint is available. Its sample request points to Subquadratic's own API, so developers should not mistake that example for Baseten endpoint documentation.

## SubQ API pricing and output format

Subquadratic's [Snap 1.1 release page](https://subq.ai/subq-snap-1-1) lists its direct API price at $0.04 per million input tokens, with no output-generation charge. That price belongs to the direct SubQ service. Dedicated infrastructure on Baseten needs a separate commercial quote.

The [Decisions API documentation](https://docs.subq.ai/endpoints/decisions-api/) labels Snap as a public preview. Requests contain the material to evaluate and named questions with criteria. A question can return a yes/no probability, choose among supplied categories or score an ordered set of levels. The response reports zero output tokens.

For a support-ticket workflow, for example, the application might ask whether a ticket needs escalation and which queue should receive it. The value comes from obtaining a bounded decision that can be checked, rather than generating prose and then extracting a decision from it.

## What the benchmark does and does not show

[Artificial Analysis](https://artificialanalysis.ai/decision-models/subq-snap-1-1) records a Decision Index score of 65 for Snap 1.1. Its index combines knowledge, reasoning, tool-use and guardrail evaluations. It is a useful comparison for decision tasks, not proof of equivalent performance on an arbitrary enterprise workflow.

The same page reports 190 milliseconds median task latency in its evaluation. That is not a measured latency guarantee for a full 384K-token request or for a new Baseten deployment.

Teams considering the model should evaluate their actual routing errors, including the cost of sending a difficult case to the wrong place. A confidence field is only useful when its behaviour has been checked on the data and decisions the application will encounter.

The provider expansion gives teams another way to operationalize a specialized model. The important comparison is then broader than token price: workload volume, hardware utilization, response time and the quality of the decisions all determine whether moving routine judgments out of a larger model is worthwhile.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders explain model launches, deployment choices and AI product updates through clear stories that developers can use.

[Book a 30-minute call](https://cal.com/paras-madan/30min).