SubQ Snap claims 61.7 on Decision Index, Jev 57.9

Alexander Whedon launched SubQ Snap, which he describes as the first long-context decision model for enterprise work. The headline claim is 61.7 on Decision Index v0.2.1 against Jev's 57.9, with Snap outperforming on 13 of 16 enterprise benchmarks. It is available today, and the first 10,000 sign-ups get 100M free input tokens.

Alexander Whedon

@alex_whedon

X

61.7 on Decision Index v0.2.1 vs Jev’s 57.9, outperforming on 13 of 16 enterprise benchmarks.

The context window is the pitch. Snap takes 384K tokens of multimodal context, which Whedon says is 12x larger than Jev, and he claims compute per token stays constant as context grows. Pricing is $0.05 per million input tokens with output free. Published latency figures scale with how much you stuff in, 80ms for minimal context, 250ms at 32k, 660ms at 100k, and 2.5 seconds at the full 384k. Whedon also says latency will be variable while the platform is tuned, and that he expects to hit capacity fast and add users as compute comes online.

What the early access testers built

Moritz Kremb had early access and built a Chrome extension that reads both text and images on a page and takes control of the browser. He describes talking to the browser, describing what he is seeing, and having it click things for him, hands-free and near instant. No latency or cost numbers from that build.

Matthew Berman ran an evolution simulation on Snap where a population sees its current form plus three random mutations every ten simulated years, across eight traits including memory, autonomy, a body, and control over its own off switch. He reports over 3000 mutations across 10 runs and more than 2000 decisions at roughly 0.68 seconds each, with the whole thing costing 21 cents. His result, which is a toy scenario rather than a benchmark, was that every simulated AI removed its off switch at some point and 9 of 10 gave it back.

What is missing

Every figure above originates with the vendor or with testers the vendor granted early access. Nobody independent has published a run of Snap against Jev on the same task, and Decision Index v0.2.1 scores are being reported by the company whose model wins them. The constant-compute-per-token claim is the interesting one to verify, because it is what would make 384K context cheap rather than merely possible. For people building on it, the 2.5 second figure at full context is the number to plan around, not the 80ms.

Matthew Berman

@TheMattBerman

X

I ran over 3000 mutations in 10 runs and it made more than 2000 decisions (~0.68 seconds). the whole thing cost 21 cents.