SubQ Snap 1.2 replaces 1.1; Artificial Analysis v2.1 reads 72.58, second among decision models

Last updated:

Source · Model card: SubQ Snap

On 11 Nov 2026 Subquadratic published SubQ Snap 1.2. The models page lists subq-snap-1.2 as the default decision model and subq-snap-latest as an alias for it. subq-snap-1.1 and subq-snap-preview were retired the same day and now route to 1.2. The documented price is still $0.04 per 1M input tokens, output free. Input is text and images. The decisions endpoint still takes noul, choice and score. Weights stay closed.

Independent board. The Artificial Analysis Decision Index page, titled v2.1, was read again on 12 Nov 2026. Among rows with kind: decisionModel, Snap 1.2 has index 72.58, time 162 ms per task, second behind Drex Max v1.1 at 72.62. Jev 1.14 on the same payload is 66.23. Gemini Decision 1 is still 65.98. These figures match the 11 Nov reading of the same page. Language-model rows on that page (Claude Haiku 5.5, GPT-6 Luna) are a different kind and are not ranked here.

The post says Snap 1.2 “scores 73, tied for first place” on that index. The page we read does not show 73 and does not show a tie with Drex Max v1.1. That sentence stays a vendor claim (UnverifiedClaim).

Other vendor claims in the same post (UnverifiedClaim). Near-perfect invariance to option order. Higher scores on LongDecisionBench, a nine-benchmark set Subquadratic is releasing. A public-component score of 69.05 on the Hugging Face Decision Index, above Snap 1.1, Jev 1.14 and d1-ultra. Snap does not appear in the Decision Index index.json read on 12 Nov (generated 11 Nov, 297 entries), so 69.05 is not a board row.

Catalog page: SubQ Snap. Capability stays 7.5. Accuracy stays 9 under the 24 Oct decision: Artificial Analysis above Jev, plus the JevBench preview row. The preview row is not a 1.2 measurement, and this round does not treat 1.2 as if it had two boards of its own.