Subquadratic Launches SubQ Snap, a 384K-Token Decision Model

Subquadratic launched SubQ Snap on 12 October 2026, a multimodal decision model that the company says handles up to 384K tokens of context, 12x larger than TypeSafe’s Jev, according to its Business Wire release and launch post the same day. Subquadratic says Snap runs on its Subquadratic Sparse Attention (SSA) and has constant compute cost per token.

On Subquadratic’s own evaluations, Snap scores 61.7 on the Decision Index public suite against 57.9 for Jev, and averages 67.6 against 61.6 across 16 enterprise-focused Decision Index benchmarks. The launch post prices it at $0.05 per million input tokens, with no charge for output tokens.

What a Decision Model Does

Subquadratic describes decision models as optimised to choose from a predefined set of options, such as which tool to use next, whether a result passes, or which step should be handed to a larger model. It says agents make these decisions repeatedly through a workflow, so moving them to a fast, specialised model can reduce the cost and latency of the whole system.

The launch post argues that context is the constraint. It says “For a decision model, nearly all inference compute is prefill”, and that under dense self-attention that cost grows quadratically with context length, so doubling the context quadruples attention compute. It says today’s decision models stay at 32K tokens.

Co-founder and CEO Alex Whedon said in the release: “Decision models get exciting when they can take real work away from LLMs, but the context window size of today’s decision models are simply too small or slow to handle real enterprise work.”

What Subquadratic Says Snap Can Do

According to the release, Snap takes text and images in the same context, up to 384K tokens, and can process 100K tokens and return a decision in about one second. The launch post says SSA reduces attention compute cost from quadratic to linear complexity, while a decision head maps the model’s representations directly to a probability distribution over the allowed answers instead of generating text.

The post names four roles Snap is designed to handle inside an agent loop: routing, tool picking, checking and stopping.

The Benchmarks, All Company-Reported

Subquadratic says Snap scores 61.7 on Decision Index 0.2.1, where every request is under about 34K tokens, against 57.9 for Jev, and leads in four of the five subareas, including arts & human taste (+12.11), language understanding (+7.46), retrieval & classification (+6.27) and tools & automation (+3.39).

Across 16 Decision Index benchmarks it maps to enterprise workflows, the company says Snap averages 67.6 against Jev’s 61.6 and leads on 13 of 16. Its table shows the largest gain on PhishNChips phishing detection (+22.0) and Snap trailing on When2Call (−7.6), API-Bank (−1.6) and ACOS (−0.6).

On long context, Subquadratic reports Decision RULER scores between 96.3 and 99.8 from 32K to 1M tokens, and 74.0 on Decision LCR with Snap reading the documents in one shot, against 48.0 for Jev on chunked input. The post says Decision LCR scores can’t be compared with published AA-LCR results, which grade written answers rather than decisions. It also reports 82.9% accuracy on Bespoke Labs’ Nimble benchmark, against 76.0% for Jev.

Price, Access and the SubQ Family

The launch post says Snap is priced at $0.05 per million input tokens, with no charge for output tokens, and is in Public Preview on a first come, first served basis while capacity is limited, with a waitlist once capacity is reached. The release says it is available today for limited access via subq.ai.

The release says Snap is part of the SubQ model family, and that in research testing at 12 million tokens the architecture cut attention compute by almost 1,000x compared with other frontier models. The post says SubQ 1.1 Small, its language model, is in Early Access with selected design partners.

The model Subquadratic measures itself against comes from TypeSafe, which raised $870M at a $7.5B valuation this month; we covered TypeSafe’s own cost and speed claims for Jev on 2 October.

The Structural Read

As we read it, the pitch is about input cost rather than output cost: the launch post says nearly all inference compute for a decision model is prefill, and the price it gives charges only for input tokens.

As we read it, the comparison Subquadratic chose matters: every headline score sets Snap against Jev, and the company reports the Jev scores itself.

As we read it, context length is the variable the company is selling: its post says today’s decision models stay at 32K tokens, against the 384K it gives for Snap.

Three Implications

AGENT BUILDERS Subquadratic says Snap is designed for routing, tool picking, checking and stopping decisions inside an agent loop.

LONG-DOCUMENT WORKFLOWS The company says Snap can reason over long codebases, documents and interaction histories in a single call, without truncation or chunking.

BUYERS COMPARING MODELS The scores Subquadratic gives are its own evaluations, and its launch post says Snap is in a Public Preview limited by capacity.

The Business Engineer Lens

This story maps onto the Business Engineer framework The State of the Inference Economy.

The framework puts it this way: “Inference is a revenue engine: continuous, distributed across millions of applications and agents, incurring marginal cost with every query.”

As we read it, decision models are a bet on that marginal cost: Subquadratic says an agent makes these calls at nearly every turn, and its launch post prices Snap at $0.05 per million input tokens with no charge for output.

What Is Not Established

Every score here is from Subquadratic’s own evaluations, including the Jev scores it reports; we have not seen independent results. The Decision LCR comparison sets Snap reading whole documents against Jev on chunked input. The two texts we read give no customer names or usage figures, and the 1,000x attention figure is from the company’s research testing of the SubQ architecture, not of Snap in production.

The Bottom Line

Subquadratic has released SubQ Snap in limited preview, a decision model it says reads up to 384K tokens at $0.05 per million input tokens and scores 61.7 to Jev’s 57.9 on the Decision Index, on its own benchmarks.