# SubQ Snap brings 384K context to structured AI decisions > SubQ Snap offers typed decisions over up to 384K tokens of text and images in public preview. Its low published token price and benchmark gains are company claims that still need independent task-level testing. By BIG CHANGE Editorial Published: 2026-10-13T07:42:19.012Z Updated: 2026-10-13T07:42:19.012Z Canonical: https://bigchange.ai/blog/subq-snap-structured-decisions-384k-preview ![Conceptual charcoal illustration of a long film strip with image and text-like frames passing across a light table beside a three-position selector switch.](https://bigchange.ai/api/media/file/subq-snap-long-context-decision-hero-v1.png) AI-generated conceptual illustration by BIG CHANGE; no actual SubQ device, interface or tested decision is depicted. Subquadratic released [SubQ Snap](https://subq.ai/introducing-subq-snap) on October 12 as a model for bounded decisions inside AI agents. Instead of writing an open-ended answer, it evaluates supplied text and optional images against questions defined by the application. Its public preview API returns choices, scores and probabilities that software can use directly. For developers, the practical change is a decision call that can take a long input without making the agent generate prose for every routing or checking step. SubQ advertises a 393,216-token context window (384K) and a price of $0.05 per million input tokens, with no output-token charge. Those are the company's current preview terms. Access is first come, first served while capacity is limited, and later sign-ups may be waitlisted. ## The big change - **What changed:** Snap offers an application-defined answer space with up to 393,216 tokens (384K) of text and image context in its public preview. A developer can ask it to select among named options or score a result without requesting generated prose. - **Why it matters:** Agent builders repeatedly make bounded calls such as tool selection and result checks. A typed response can go straight into the next software step, while a generative model remains available for work that needs an open-ended answer. Whether that division improves a real agent depends on decision accuracy and the surrounding workflow. - **What to watch:** SubQ's launch benchmarks favor Snap on several decision tests, but they are company-run evaluations. Independent, task-level tests and broader preview access would show how its speed, cost and error rates behave in deployed agent workflows. ## What a Snap call returns The [Decisions API documentation](https://docs.subq.ai/endpoints/decisions-api/) names `subq-snap-preview` as the model for `POST /v1/systemone`. A request supplies a `state`, which can be text or structured data, and one or more named questions. A *noul* is a yes-or-no judgment expressed as a probability of yes. A *choice* selects among two to 255 named options and returns probabilities for each. A *score* uses two to ten ordered levels and returns a weighted score with its probability distribution. Those shapes matter more than the product label. The developer defines the allowed answers and decides what the application does with them. The API's probabilities are model outputs, not verified measures of how often a particular real-world decision will be correct. SubQ's own documentation shows developers applying thresholds and sending uncertain cases back to a person; it does not establish a universal threshold for every task. Snap accepts up to four PNG, JPEG or WebP images alongside the state. Each image may be at most 4 MiB decoded, and their combined decoded size may be at most 8 MiB. The API documentation says image tokens and text count toward the same 393,216-token context limit. Video is unsupported; its suggested route is to sample no more than four frames as images. These are concrete input limits for the preview API, rather than a promise that the model will interpret every long document or image reliably. ## The benchmark gap to keep in view In its [launch evaluation](https://subq.ai/introducing-subq-snap), SubQ reports a 61.7 score for Snap against 57.9 for Jev on Decision Index 0.2.1's full public suite. It also reports a 67.6 average against 61.6 across a selected group of 16 enterprise-oriented benchmarks. Snap leads on 13 of those 16 and trails on three, including a test of when an agent should call, ask, answer or decline. These figures come from SubQ's evaluation; BIG CHANGE has not reproduced them or tested either model in an agent. The company also reports a 96.3 result for Snap on its adapted Decision RULER test at one million tokens. That result belongs to a separate benchmark evaluation. The [current API request limit](https://docs.subq.ai/endpoints/decisions-api/) is 393,216 tokens, and the documentation says an over-limit request fails. Readers should not use the one-million-token benchmark as the preview API's input specification. SubQ says Snap can process a 100,000-token request and return a decision in about a second. The public claim does not give BIG CHANGE an observed latency distribution across accounts, concurrent traffic or different question types. The preview should be assessed against the reader's own decision set, error costs and response-time requirements before it replaces a step in a consequential workflow. ## Access, price and data handling The [model listing](https://docs.subq.ai/using-the-api/models/) marks Snap as a public preview with text input and image input; it lists $0.05 per million input tokens and free output. The launch page says capacity may put new users on a waitlist. Neither source establishes that every organization has immediate access or a particular rate limit. Teams evaluating sensitive work should also check the [data-retention policy](https://docs.subq.ai/using-the-api/data-retention/) before sending documents or images. The documented organization default retains prompts, completions and conversation history until deletion. SubQ offers a fixed-day policy and organization-wide zero data retention on request. It says billing and support records retain model, token counts, timing and metadata tags even under zero data retention, without prompt or completion content. The Decisions endpoint remains available under that organization setting. BIG CHANGE has not verified the service's storage behavior independently. SubQ's contribution is a specific API shape for repeated bounded decisions over large context, at a published low token rate. The unresolved question is how often those decisions hold up on the work an organization would actually route through it. That answer requires independent testing beyond launch scores. ## Sources & further reading - [SubQ's October 12 launch post](https://subq.ai/introducing-subq-snap) explains the intended agent role, pricing, preview access and company-run evaluations. Its speed and benchmark figures are vendor claims; the one-million-token Decision RULER result is separate from the preview API cap. - [SubQ Decisions API reference](https://docs.subq.ai/endpoints/decisions-api/) documents the endpoint, typed questions and responses, token and image limits, and error behavior. It is current product documentation, not evidence of reliability in a deployed workflow. - [SubQ model listing](https://docs.subq.ai/using-the-api/models/) gives the preview model ID, modality and published price. Terms and availability can change during preview. - [SubQ data-retention documentation](https://docs.subq.ai/using-the-api/data-retention/) describes the default and optional organization policies. The policy describes intended service behavior; BIG CHANGE did not audit storage or deletion. ## Sources - [SubQ: Introducing SubQ Snap](https://subq.ai/introducing-subq-snap) — October 12, 2026 official launch post. Establishes intended agent role, preview availability, pricing and company-run benchmarks. Speed and benchmark figures are vendor claims, not independently replicated outcomes. - [SubQ Decisions API](https://docs.subq.ai/endpoints/decisions-api/) — Current official endpoint specification, read October 13, 2026. Documents typed questions and responses, token and image limits, and errors. It does not establish reliability in deployed work. - [SubQ Models](https://docs.subq.ai/using-the-api/models/) — Current official model listing, read October 13, 2026. Gives preview model ID, modalities and US-dollar token rates. Terms may change. - [SubQ Data retention](https://docs.subq.ai/using-the-api/data-retention/) — Current company data policy, read October 13, 2026. Describes default and optional retention settings. BIG CHANGE did not audit storage behavior.