What SubQ Snap is
SubQ Snap is a System One model from Subquadratic, a research company that builds long-context models on its own sparse attention method. The first release came out on 12 October 2026 and SubQ Snap 1.1 on 22 October. You send text, structured data or up to four images with a set of typed questions, and it returns a probability for each outcome. It is a hosted API with closed weights, in public preview.
What it returns
The endpoint is POST https://api.subq.ai/v1/systemone, with subq-snap-1.1 as the model. The request has Jev’s state and named questions. A Noul returns the probability of yes. A Choice returns a probability per option and a confidence. A Score returns a probability per level and the probability-weighted level. SubQ’s docs do not say whether TypeSafe’s SDKs work against it, and they list no SubQ SDK, so the examples call the endpoint with fetch or requests.
What it is good at
Its main difference is context. One request can hold 384K tokens, against 64,000 for Jev, so a whole contract or a long agent history fits in one call. Subquadratic says Snap answers a 100,000-token request in about a second. That is a vendor figure. Subquadratic’s examples include picking files and tools inside coding agents, a form of intent and model routing, and support inbox triage.
On the Artificial Analysis Decision Index, checked on 23 October 2026, SubQ Snap 1.1 scores 65.0. That is second of 17 decision models, 0.25 behind d1-ultra and just ahead of Elephant 9B at 64.7. Subquadratic’s site calls it tied for first. Its non-hallucination rate on AA’s knowledge questions is 85.7, second of the 17, so it often answers “I don’t know” rather than guess. Its When2Call tool-calling accuracy is 94.6%, also second. Its calibration error, which this site computed from AA’s published bins, is 0.082. AA measured $30 per million tasks and a median of 190 ms per task.
Benchmark Heaven’s JevBench measured the first release on 13 October 2026 on its hosted API board: a Capability Score of 81.5, against 75.0 for Jev 1.14. It has not yet measured 1.1.
Subquadratic’s own runs give 1.1 a chance-corrected 65.7 on the Jev Decision Index 0.2.1 public suite, against 60.6 for Jev 1.14, and 72.6 against 50.7 on its own long-context test. These are vendor-run. The multimodalart Jev Decision Index has no row for SubQ Snap.
What it is not for
It writes no text, so it cannot explain an answer. Its accuracy on AA’s knowledge questions is low, at 20.2%, so it is a poor fit for questions that depend on recalled facts. Subquadratic’s own tables put it 22 points behind Jev 1.14 on PhishNChips phishing detection, which it blames on its default threshold. The rate limit is not published.
Access
Sign up at console.subq.ai, buy credit, and call the API with Authorization: Bearer. It is in public preview, which SubQ says is open to every account and can change before general availability. Inputs and outputs are deleted within 60 days, and zero data retention is available.
Specifications
| System One Score | 97.7 from 2 of 3 boardsArtificial Analysis Decision Index 99.7, JevBench 95.7. How the score works |
| Question types | ChoiceScoreNoul |
| Max Choice options | 255 |
| Score levels | Up to 10 |
| Questions per call | 128 |
| Total context | 393,216 tokens |
| State budget | Not documented |
| Rate limit | Not published as of 23 October 2026. Going over the limit returns HTTP 429. |
| Endpoint | POST https://api.subq.ai/v1/systemone |
| Licence | Not published as of 23 Oct 2026 |
| SDKs |
The docs give a 384K-token (393,216) context window for the whole request; a request over the limit, or close to it, fails with 400. A request takes 1 to 128 questions, a Choice 2 to 255 options and a Score 2 to 10 levels. Up to 4 PNG, JPEG or WebP images go in an images array as plain base64, at most 4 MiB each and 8 MiB in total, and each uses about 64 to 256 tokens. The body can be up to 13 MiB. Video is not supported.
Versions
- subq-snap-1.1, 22 Oct 2026, The current model and the default decision model in SubQ's docs, in public preview. Released with the launch post of 22 October 2026 at $0.04 per million input tokens. Closed weights. Release notes
- subq-snap-preview, 12 Oct 2026, The first SubQ Snap, launched in public preview on 12 October 2026 at $0.05 per million input tokens, with a waitlist once capacity ran out. SubQ's models docs list only subq-snap-1.1 as of 23 October 2026. Release notes
Use cases
What people use SubQ Snap for, one page per pattern.
Intent and model routing with System One models
One Jev call reads an incoming request and returns its intent as a label plus a difficulty rating on a scale you wrote. Your router reads both numbers and picks the handler: deterministic code, a cheap model, an expensive one, or a human queue.
Support inbox triage with System One models
Send a support ticket to Jev once with every question attached. Category comes back as a selected label, severity and frustration as numbers on scales you wrote, refund intent as a probability. Your code reads those values and decides what happens to the ticket.
Benchmarks
Published scores, each linked to its source. A row marked vendor-run was measured by the vendor itself, so read it as one side's result. The compare hub sets these beside the other models.
| Benchmark | Metric | Score | Run by | Date |
|---|---|---|---|---|
| Artificial Analysis Decision Index 1.0 | Index score | 65 score | Independent | 23 Oct 2026 |
| Artificial Analysis Decision Index 1.0 | Expected calibration error | 0.0819 ECE | Independent | 23 Oct 2026 |
| Artificial Analysis Decision Index 1.0 | Cost per 1M tasks | 30 USD | Independent | 23 Oct 2026 |
| Artificial Analysis Decision Index 1.0 | Time per task | 190 ms | Independent | 23 Oct 2026 |
| Decision LCR 1.1 (Subquadratic) | Accuracy | 72.6 % | vendor-run | 22 Oct 2026 |
| Jev Decision Index 0.2.1 public suite (Subquadratic run) | Skill score | 65.7 score | vendor-run | 22 Oct 2026 |
| JevBench 1.7.47 | Capability Score | 81.5 score | Independent | 13 Oct 2026 |
- Artificial Analysis Decision Index, Index score: Row: SubQ Snap 1.1, hosted by Subquadratic, 2nd of 17 decision models as of 2026-10-23, 0.25 behind d1-ultra. Sub-scores: Omniscience accuracy 20.2, non-hallucination 85.7, HLE-MultiChoice 17.5, When2Call 94.6, WildGuardTest F1 94.9. The row was missing from one copy of the page fetched on 2026-10-23 and back in the next.
- Artificial Analysis Decision Index, Expected calibration error: Row: SubQ Snap 1.1. Computed by this site from AA's published calibration bins (all answers pooled); AA does not publish an ECE. Lower is better.
- Artificial Analysis Decision Index, Cost per 1M tasks: Row: SubQ Snap 1.1. AA's measured cost at list prices on its four evals.
- Artificial Analysis Decision Index, Time per task: Row: SubQ Snap 1.1. Median time per task.
- Decision LCR 1.1 (Subquadratic), Accuracy: Subquadratic's long-context decision test, Snap 1.1 reading each request in one pass against Jev 1.14 at 50.7 with the input split into chunks. Vendor-run.
- Jev Decision Index 0.2.1 public suite (Subquadratic run), Skill score: Run by Subquadratic on SubQ Snap 1.1, chance-corrected. It gives Jev 1.14 60.6 and the first SubQ Snap 61.8. The 12 October launch post gave the first release 61.7 against 57.9 for Jev.
- JevBench, Capability Score: Row: SubQ Snap (preview), model subq-snap-preview, measured 2026-10-13 on the first release, not 1.1. Intelligence 72.3, Calibration 90.7. Fourth by Capability Score on the hosted API board.