ENPT

SubQ Snap

Last updated:

Beta80–2500ms$0.04/M input

Quadrant scores

See the full quadrant

Scored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.

  • Maturity5.3/10

    Public preview hosted API (subq-snap-preview), first come first served while capacity lasts; vendor warns latency will vary. Decisions API docs. Closed weights, no SLA yet.

  • Capability7.5/10

    Official JevBench (17 Oct): Intelligence 72.3, calibration 90.7, p50 0.47 s. Noul / choice / score over up to 384K tokens of text and images, on Subquadratic Sparse Attention plus a decision head. No ECE. Snap 1.1: AA DI 65 (tied #1). Accuracy 9 counts AA plus the JevBench preview.

  • Adoption4/100

    Launch post ~3k likes and 4.4M views (13 Oct); 100M free tokens for the first 10,000 sign-ups. No named production users yet.

Vendor claims

News covering this model

publicado (OK do Fabricio 17/10 BRT). Board audit of 17 Oct 2026: see the “Board audit” section below; where it conflicts with older board statements on this page, it wins.

SubQ Snap 1.1 (22 Oct 2026)

publicado (OK do Fabricio 24/10 BRT). Prices and board status below this section are superseded where they conflict.

Subquadratic released SubQ Snap 1.1 on 22 October 2026 (post; @alex_whedon, 16:48 BRT). Model id subq-snap-1.1 on POST https://api.subq.ai/v1/systemone; still closed weights. Price drops 20% to $0.04 per 1M input tokens, output free; context stays at 384K tokens.

Independent board. Artificial Analysis puts Snap 1.1 at 65 on its Decision Index v1.0, tied for first with d1-ultra and Elephant 9B (mammoth-1 family), 8 points ahead of Jev 1.14 (57). AA attributes the result to a low hallucination rate: non-hallucination 86% (Elephant 9B 93%, d1-ultra 85%), accuracy 20% on its knowledge questions, 18% on HLE-MultiChoice, $30 per 1M tasks. Snap 1.1 is not on the Decision Index or JevBench yet (JevBench still shows the preview).

Subquadratic’s numbers (UnverifiedClaim): +7.2 points over Jev 1.14 across 16 enterprise benchmarks (+5.5 over Snap 1.0); When2Call 75.2 → 96.2; expected calibration error 0.108 → ~0.014 on held-out problem families; confident hallucinations on out-of-distribution queries 99.3% → 12.5%.

Update 24 Oct: the Snap 1.1 post also lists HF Decision Index 0.2.1: 65.7 (Jev 1.14 60.6, Snap 1.0 61.8). That is Subquadratic’s own run on an old edition, not a board row (UnverifiedClaim). The only independent evidence for 1.1 is still the Artificial Analysis 65 above. Score applied: capability 7.5.

Score (approved 24 Oct): capability 7.3 → 7.5 = 9 × 0.30 + 5 × 0.25 + 7 × 0.25 + 9 × 0.20 = 7.50. Accuracy 9: two sources above Jev, Artificial Analysis 65 (1.1) and the JevBench preview row 72.3 (63.6 for Jev), counted together by Fabricio’s decision. Maturity 5.3 and adoption 4 unchanged. Working in docs/06-quadrante.md.

Read this first. Every number below is Subquadratic’s (UnverifiedClaim). The 61.7 is on Decision Index 0.2.1, an older edition than the current 0.3 board, so it can’t be ranked against 0.3 scores. Snap is not on the 0.3 board.

SubQ Snap (launch post, 12 Oct 2026) is a hosted decision model built on Subquadratic Sparse Attention (SSA), which the company says keeps compute per token constant as context grows. A decision head returns probabilities over the allowed answers, so nothing is generated. It accepts text and images in the same context, up to 384K tokens, about 12× Jev’s window by the vendor’s count. Subquadratic pitches it for routing, tool picking, checking and stopping inside agent loops where the evidence is long: a codebase, a contract and its exhibits, a long tool history.

Price: $0.05 per million input tokens; output is free. Availability: public preview, first come first served; new users may be waitlisted when capacity runs out.

Vendor results (UnverifiedClaim)

Benchmark Snap Jev
Decision Index 0.2.1 61.7 57.9
16 enterprise DI benchmarks (average) 67.6 61.6
Decision LCR (72K–115K, one-shot vs chunked) 74.0 48.0
Bespoke Nimble Suite 13 82.9% 76.0%

Snap loses to Jev on When2Call (−7.6), API-Bank (−1.6) and ACOS (−0.6). Decision RULER and Decision LCR are Subquadratic’s own conversions of RULER and AA-LCR to multiple choice, not public boards.

Score working (13 Oct 2026)

  • Maturity = 6 × 0.30 + 7 × 0.25 + 3 × 0.25 + 5 × 0.20 = 1.80 + 1.75 + 0.75 + 1.00 = 5.30 → 5.3
  • Capability = 9 × 0.30 + 5 × 0.25 + 7 × 0.25 + 7 × 0.20 = 2.70 + 1.25 + 1.75 + 1.40 = 7.10 → 7.1
  • Adoption = 9 × 0.35 + 2 × 0.35 + 1 × 0.30 = 3.15 + 0.70 + 0.30 = 4.15 → 4

Calibration 5: no ECE or reliability data. Latency 7: detailed vendor figures by context length. Accuracy 7: a vendor DI run with a published method, counted as evidence per Fabricio’s rule but still unverified and on the old 0.2.1 edition.

Board audit (17 Oct 2026)

publicado (OK do Fabricio 17/10 BRT). Checked against the official boards on 17 Oct 2026: Decision Index 0.3.2 (data of 16 Oct, 19:00 UTC), JevBench and ImageJevBench live. The boards are separate sources and are never averaged. Older board statements and rankings further down this page are superseded where they conflict with this section.

Board Row Result
JevBench API (live 17 Oct) SubQ Snap Capability 81.5; Intelligence 72.3; calibration 90.7; p50 0.47 s; $0.065/1k

Board-status statements on this page to recheck against the table (may be superseded):

  • “The 61.7 is on Decision Index 0.2.1, an older edition than the current 0.3 board, so it can’t be ranked against 0.3 scores.”

Score update (audit, 17 Oct 2026). Capability 7.1 → 7.3 = types × 0.30 + calibration × 0.25 + latency × 0.25 + accuracy × 0.20 = 9 × 0.30 + 5 × 0.25 + 7 × 0.25 + 8 × 0.20 = 7.30 → 7.3. Types, calibration and latency come from the earlier working (9 / 5 / 7); only accuracy is re-derived from the board (8; was 7): JevBench Intelligence 72.3 ÷ 63.6 × 8 = 9.1, half-up, capped at 8. Maturity and adoption unchanged.