Drex 1.5 vs Decision-1: The Models That Score, Not Write

Drex 1.5 is the decision model Nace.AI released around October 9, with 8.95 billion parameters and open weights (Nace.AI Open RAIL-M license). It generates no text: given a state (text or JSON) plus typed questions (choice, yes/no, ordinal score), a single forward pass returns a calibrated probability for each option via POST /v1/systemone, compatible with TypeSafe's Jev decision API. Default context is 16,384 tokens, expandable to 131,072; bf16 weights are about 18GB, runnable on one CUDA GPU, and a 9.5GB Q8_0 GGUF build runs on Apple silicon or CPU.

Decision-1 is the decision-scoring model Microsoft announced October 9 on its Command Line blog, post-trained from Alibaba's Qwen3.5-9B; Microsoft says it will later rebase it on its own models and OpenAI's. It is hosted-API only (Microsoft Foundry and OpenRouter) with no open weights; context is 32,768 tokens, input $0.042 per million tokens, output free. Microsoft reports the highest average accuracy, 83.5%, in a 36-benchmark comparison (nearly 150,000 questions kept blind from training), with 85ms P50 latency — 2.5x faster than H2O-Lightning-4B and 35x faster than GPT-6 Sol — all from Microsoft's internal testing.

The table below compares public pricing, context, openness and vendor-reported benchmarks; note the two cite different benchmark suites, so the scores are not directly convertible.

1. Launch: : Drex 1.5 went live around October 9 (OpenRouter lists 2026-10-09); Microsoft announced Decision-1 on October 9 on its Command Line blog.
2. Pricing: : Drex 1.5 is $0.04 input / $0 output per million tokens on OpenRouter; Decision-1 is $0.042 input per million tokens with free output (Microsoft).
3. Benchmarks: : Drex 1.5 scores 58.08 on Decision Index 0.3.1 (Nace model-card data, highest under 10B parameters); Decision-1 averages 83.5% accuracy in Microsoft's 36-benchmark comparison (Microsoft internal testing). The suites differ and are not directly comparable.
4. Openness: : Drex 1.5 has open weights (RAIL-M license) and supports local deployment; Decision-1 is hosted-API only, with no open weights or local build.