The decision-model price war: SubQ Snap, Microsoft-Decision-1, Liquid d1 on MLX, SemIf on SageMaker
Sources: Gmail (Subquadratic launch email, AI Breakfast 10-12), lab watcher (Liquid AI HF repos, AWS ML blog), X Following feed. Raw: raw/gmail/2026-10-13-newsletters.md, raw/labs/2026-10-13-000834-labs.md, raw/twitter/feed/2026-10-1[23]-*-ranked.json.
TL;DR
A decision model reads a state (text, JSON, image) plus a fixed set of options and returns a calibrated probability per option in one forward pass, generating no text. It is the cheap gate in front of a frontier model. In one US day the category got prices, a cloud deployment path, and on-device builds:
- SubQ Snap (Subquadratic): 384K multimodal context, decision in under a second on 100K tokens (660 ms at 100K, ~2.5 s at 384K, roughly linear, consistent with its subquadratic attention), $0.05 per million input tokens, output free. Claims 61.7 on Decision Index v0.2.1 vs 57.9 for Jev, winning 13 of 16 enterprise benchmarks. Vendor numbers.
- Microsoft-Decision-1: post-trained from Qwen3.5-9B, $0.042 per million input, free output, 32K context, 85 ms p50 / 125 ms p95. 83.5% across 36 benchmarks vs 81.9% for Quyet-1.0-Large, but calibration 92.2 vs Quyet's 93.1. H2O.ai disputes the competitor latency figures.
- Liquid d1-3B and d1-omni-600M on MLX (fp16, 8-bit, 4-bit). Several named, typed questions (yes/no/null, choice, score) answered over one state in one pass. The card reports 8-bit moves probabilities by ~0.015 but 4-bit by up to ~0.2, a warning that calibrated outputs are more quantization-sensitive than accuracy.
- AWS: SemIf on llama.cpp in SageMaker (AWS ML blog). SemIf (formerly OpenJev) reads final-position logits to return probabilities over declared options without decoding; AWS ships a child image of its llama.cpp container for CPU or GPU. The first hyperscaler deployment recipe for the category.
- Practitioner test: one builder put Jev inside an open agent harness and reports Opus-level accuracy on 40 decisions at ~150x lower decision cost, with low-confidence answers routed to review (X article). Small sample.
The gate now has a price per million tokens
Output is free because there is no output: one forward pass, one probability per option.
flowchart LR
S["State<br/><small>ticket, doc, screenshot</small>"] --> G["Decision model<br/><small>one pass, no text</small>"]
G -->|p above threshold| A["Act<br/><small>route, label, tool</small>"]
G -->|uncertain| F["Frontier LM<br/><small>reasons on hard cases</small>"]
F --> A
classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
classDef loop fill:#fff3bf,stroke:#f08c00,color:#1b1b1b,stroke-width:2px
classDef core fill:#e5dbff,stroke:#6741d9,color:#1b1b1b,stroke-width:2px
classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
class S input
class G loop
class F core
class A exit
linkStyle 1 stroke:#2f9e44,stroke-width:2px
Blue is the input state, amber is the cheap gate, purple is the frontier model, green is the action.
Relation to prior wiki pages
- Continues Decision 3.0 / THX-01 (10-12), where the gate learned to read screenshots and THX-01 published an ECE of 0.003. Today the category moved from capability to price: two vendors at ~$0.04-0.05/M input with free output. That is ~100x below frontier input prices.
- The 10-12 Looking Ahead asked for an independent calibration audit within 60 days. Microsoft's own table already shows the accuracy leader losing on calibration, and Liquid's card shows 4-bit quantization degrading calibration; the audit question is getting sharper, not resolved.
- Production echo: Meta's ads system (efficiency shorts 10-13) routes a variable share of traffic to a bigger model based on spare capacity, the same gate-then-escalate shape at fleet scale. AMD's Helion rack (10-13 hardware page) pitches a built-in token router.
Gaps
All benchmarks are vendor-run on vendor-defined indexes (Decision Index, JevBench). No shared calibration benchmark yet. Latency claims are not measured under the same hardware.