TLDRocket
10 October 2026
Today’s biggest AI thread is a quiet but consequential shift away from “write the answer” toward “score the options.” Microsoft’s Microsoft-Decision-1, built on a Qwen3.5-9B core, returns calibrated probabilities over fixed choices for routing, classification, verification, and agent control—no free-form prose. Microsoft says it hits 83.5% average accuracy across 36 benchmarks with 85 ms p50 latency, exposed through a closed, hosted API that can be deployed via Microsoft Foundry or OpenRouter. Pricing is $0.042 per million input tokens, with free output, and the output format itself—JSON option probabilities—signals where reliability work is heading. Nace.AI’s Drex 1.5 makes the same move from the open-source side, offering a 9B decision model that scores provided options rather than generating text (it reports 58.08 on the Decision Index 0.3.1 leaderboard).
Read the full briefing →