SubQ Snap: Subquadratic's 384K-Context Decision Model at $0.05

SubQ Snap is Subquadratic's new decision model, and its pitch is simple: read a lot more before you decide. Released October 12, it takes up to 384K tokens of text and images in one call and returns a probability for each allowed answer instead of writing text. It costs $0.05 per million input tokens, with output free, and is in public preview.

What actually happened

Subquadratic published the launch post, "Introducing SubQ Snap", on October 12. The company says it focuses on how model algorithms scale with compute, memory and data. Earlier this year it announced Subquadratic Sparse Attention (SSA), the architecture behind its SubQ language model, and the new model is built on the same research.

A decision model handles the small, bounded calls inside an agent loop. Which file matters, which tool to call next, whether a test passed, whether to escalate. You send context, a question and a set of allowed answers, and you get a probability for each answer back. Subquadratic names four jobs: routing, tool picking, checking and stopping.

Its main claim is about context length. "Today's decision models stay at 32K tokens and decide without most of what they need to see," the post says. SSA is meant to cut attention cost from quadratic to linear in context length, so long inputs stay cheap and fast.

The numbers, all company-reported

Every benchmark in the post compares SubQ Snap with a single rival, Jev. On 16 enterprise-style Decision Index tasks, Subquadratic reports an average of 67.6 vs 61.6 and wins on 13 of 16. The biggest gaps are phishing detection (47.1 vs 25.1), code output prediction (73.8 vs 57.1) and RAG hallucination detection (64.8 vs 51.3).

It loses three. When2Call, a test of whether to call a tool, ask, answer or decline, goes to Jev 74.6 to 67.0. API-Bank and an aspect-level review task are narrow losses.

On long context, Subquadratic reports 96.3 to 99.8 on its Decision RULER adaptation from 32K up to 1M tokens. On Decision LCR, built from 72K–115K token document sets, it reports 74.0 against Jev's 48.0 with chunking.

How it fits the decision model race

This is a crowded lane all of a sudden. CellCog's breakdown counts it as the sixth hosted decision model in about three weeks, after Jev, Perplexity's model, Cloudflare's Clef, OpenAI's Decisions API and Microsoft-Decision-1. We covered several of them: TypeSafe Jev, Microsoft Decision-1 and the OpenAI Decisions API.

On price it sits mid-pack. CellCog puts SubQ Snap about 19% above Jev's $0.042 per million input tokens, and says the first 10,000 users get 100M free tokens. At $0.05 per million, a full 384K-token decision costs about 1.9 cents.

CellCog also flags what is not verified. No outside leaderboard has run the model yet. The long-context tests are Subquadratic's own multiple-choice adaptations, and the post itself says its LCR scores can't be compared with published AA-LCR results. The RULER chart runs to 1M tokens, past the 384K the product accepts.

More context per decision is the pitch. Independent evals are the proof, and they are not in yet.

Why decision models matter for agents

Most agent loops spend a surprising share of their budget on tiny choices. A frontier LLM writing a paragraph just to say "use the search tool" is expensive and slow. Routing those bounded calls to a specialized model and saving the big model for real generation is the whole economic case for this category.

The catch Subquadratic is attacking is context. A three-way tool choice can still depend on a constraint buried 40 turns back or a function defined in another module. If the decision model can only see 32K tokens, someone has to chunk or summarize first, and that step can drop the one detail that mattered.

What to do as a reader

If you build agents, the realistic test is your own loop. Take a routing or tool-picking step that currently truncates context, run it through SubQ Snap with the full artifact, and compare accuracy and latency against what you use today.

Pay attention to the When2Call loss. Deciding whether to act at all is one of the four jobs Subquadratic sells, and that is the benchmark where it trails Jev. If your agent's riskiest call is "should I do this," test that case first.

If you are tracking AI tokens or hype, note that this is an API product from a private company, not a coin. Nothing here is investment advice.