SubQ Snap: long-context decisions, API setup and cost

Open the SubQ console, create an account and obtain an API key. Public Preview is first come, first served while capacity is available; a waitlist can apply once it fills. Use subq-snap-preview, not the separate SubQ language-model preview ID.

Call POST https://api.subq.ai/v1/systemone with a state and named questions. A state can be text, an object or an array. Define the answer space before calling: a choice for routing, a noul for a yes/no gate, or an ordered score for a rubric.

This request routes a damaged-delivery ticket. Read answers.route.choice and the returned probability distribution. Your application chooses when to act and when to ask for review; retain the question text and criteria with each decision so you can inspect a surprising result.

The request must fit 393,216 tokens, documented as 384K. Include the state, questions and image tokens in that budget; a request at or near the limit can fail with 400. Subquadratic also reports long-context benchmark runs at 1M tokens. Those evaluate a different context setting; plan public API requests against 384K.

Images are PNG, JPEG or WebP bytes encoded as base64 in an object with mediaType and base64 fields. Send up to four images, at most 4 MiB each and 8 MiB decoded in total. The base64 field contains the encoded bytes without a data: prefix. The docs estimate roughly 64–256 image tokens per image; inspect returned input usage for billing.

The API does not accept video. For a visual check, send up to four selected frames as images and make the timing or coverage relevant to the question explicit. A few frames suit a bounded scene check; keep a separate video workflow when the decision depends on events between them.

Subquadratic’s October 12 launch lists $0.05 per million input tokens and no output charge. A 100,000-input-token request costs $0.005. If 10,000 requests each use that much input, the input budget is $50; at 10,000 tokens per request, the same request count costs $5.

The response reports usage.input_tokens and output_tokens: 0. Multiply billed input by $0.05/1,000,000 and add the usage of repeated or retried calls. A downstream model that writes an answer or patch has its own bill. Long context can avoid a retrieval or summarization step, but every repeated context still contributes to input usage.

Subquadratic reports 61.7 for Snap against 57.9 for Jev on Decision Index v0.2.1. Its launch describes a short-context suite with requests under roughly 34K tokens, plus separate enterprise and long-context evaluations. The 1M-token Decision RULER score is 96.3; it measures retrieval decisions at that evaluation length.

The launch’s Decision LCR comparison uses Snap on a whole input and Jev with chunking. Treat the input preparation as part of that comparison. For your own routing task, keep the label set and evidence fixed and track incorrect routes, deferrals, response latency and total billed input.

Snap supplies hosted long-context decisions with text and image input. Decision 3.0 offers Apache 2.0 checkpoints and local text, image and video workflows; you operate its memory, runtime and serving capacity. Clef offers local decision checkpoints as well as hosted variants with distinct context and input support.

Compare them on the same decision, using the evidence each endpoint can actually accept. A hosted API budgets tokens and retries; self-hosting budgets provisioned compute, idle time and operations. Use the decision calculator for a supported local checkpoint after measuring its latency, rather than treating Snap’s token rate as a GPU hourly price.