The Pulse
Subquadratic Launches SubQ Snap With a 384K-Token Context
Subquadratic introduced SubQ Snap, a decision model for agent workflows with a context window of up to 384,000 tokens of text and images. The company says the model costs $0.05 per million input tokens and is available in public preview.
AI.info Team · Written with AI ·
Subquadratic introduced SubQ Snap on October 12, describing it as a decision model for agent workflows that can process up to 384,000 tokens of text and images in one request. The company says requests containing up to 100,000 tokens can return a decision in under a second. Snap is designed to select tools, route work, check results and determine when a workflow should call a larger language model.
A model for bounded choices
Unlike a chatbot asked to produce an open-ended response, Snap is built to choose among answers defined by the application. Subquadratic’s examples include deciding which files matter, selecting the next tool and checking whether a test passed. The model returns structured decisions rather than a passage of generated prose.
That format can suit repeated steps inside an automated workflow, where an agent needs a yes-or-no judgment, a choice from a set of options, or a score. Subquadratic’s API documentation describes typed answers and probabilities, with confidence values available for some response types. The company’s argument is that a specialist model can take on these bounded calls while a larger generative model handles work that requires writing or extended reasoning.
384K tokens, with a separate million-token test
Subquadratic attributes Snap’s long-context capacity to Subquadratic Sparse Attention, or SSA, its approach to reducing the attention computation required as input length grows. The company says the model accepts text and images together, including material such as screenshots, charts and scanned pages. The product limit it publishes is 384K tokens.
That limit is not the same as the longest context in the company’s benchmark results. Subquadratic says Snap scored between 96.3 and 99.8 on its Decision RULER tests from 32K to one million tokens. The company presents the one-million-token result as an evaluation, while its product description specifies a maximum context of 384K; the benchmark figure should not be read as an API promise to handle one million tokens.
On Decision LCR, a test built around multi-document inputs of 72K to 115K tokens, Subquadratic reports a score of 74.0 for Snap and 48.0 for Jev, a competing decision model. The company says the Jev result used chunking, while Snap answered in one pass. Subquadratic cautions that the scores are not directly comparable with published results for the original AA-LCR benchmark, which grades written answers rather than structured decisions.
Company benchmarks show a mixed lead
Across 16 enterprise-focused Decision Index benchmarks, Subquadratic reports an average score of 67.6 for Snap against 61.6 for Jev, with Snap ahead on 13 tasks. The evaluations cover examples including phishing detection, contract review, customer-support routing, code-output prediction and tool selection. The company’s reported gains vary by task: Snap leads Jev by 22 points on PhishNChips but trails it on two tests, including a measure of when to call, ask, answer or decline.
On the broader Decision Index 0.2.1 suite, where the company says requests stay below 34K tokens, Snap scores 61.7 against Jev’s 57.9. These are Subquadratic’s published evaluations, not an independent assessment of the model. The results point to an important distinction in its pitch: a large context window can reduce the need to trim or split inputs, but performance still depends on the decision task and the evidence provided.
Public preview, priced by input
Subquadratic lists Snap as a public-preview model, available through its Decisions endpoint. Its documentation identifies the model as subq-snap-preview, says it accepts text and image input, and describes support for up to four images per request. The company says access is limited by capacity and may move to a waitlist once that capacity is reached.
Pricing is $0.05 per million input tokens, with no charge for output tokens, according to Subquadratic. The price makes the amount of context sent to the model a direct part of the cost calculation, even when the answer itself is free. For now, the concrete offering is a preview decision API with a 384K-token ceiling—not a general-purpose replacement for language models that generate longer-form responses.