Microsoft's new decision model is 35x faster than GPT-6 Sol

If you have ever built an AI agent workflow, you know the pain: chaining calls to large language models for every single branch turns your pipeline into a sluggish, budget-draining mess.

Software rarely needs a novel when making a choice. Most of the time, all an agent needs is a fast, reliable decision—whether to proceed, retry, or call a human.

Microsoft just rolled out Microsoft-Decision-1 to take over this exact workload. Now available in Microsoft Foundry and on OpenRouter, the dedicated decision-scoring model handles routing, classification, prioritization, verification, and workflow control.

Tuned for lightning-fast choices

Microsoft built the model by post-training Qwen3.5-9B for single-pass decision scoring. Future versions are slated to run on Microsoft AI and OpenAI foundations.

Instead of generating free-form text, Microsoft-Decision-1 outputs calibrated probability scores across a fixed set of options. It handles yes/no answers, multiple-choice menus, numeric ratings, and rubric-based evaluations through a single structured API call.

Across 36 blind benchmarks covering nearly 150,000 questions, Microsoft-Decision-1 logged the top accuracy score. In latency tests, its P50 speed proved 2.5 times faster than runner-up H2O-Lightning-4B v1.1 and roughly 35 times faster than GPT-6 Sol.

Decision models often fall apart when inputs change slightly, but Microsoft tested this with eight perturbation methods. The model flipped decisions on only 1.3% of tests on average, registering zero flips when choices were reordered or paraphrased.

On safety, tests across 5,250 requests over 11 jailbreak and prompt injection benchmarks showed it consistently blocked harmful inputs without breaking legitimate workflows.

Pricing is set at rock-bottom rates: $0.042 per million input tokens, with free output tokens.

Tested inside Xbox and Copilot

Microsoft put the model through internal production tests before releasing it publicly.

Xbox Research used it to sort over 10,000 pieces of open-ended player feedback from Steam, surveys, and X into fixed research themes. The model delivered classification quality on par with GPT-6 Sol, while running over 14 times faster and 200 times cheaper.

The Copilot team tested it as an automated judge to rate chat responses, finding quality competitive with GPT5.6 Luna while operating 100 times faster.

In site reliability workflows, on-call engineers found it retrieved incident knowledge from logs and tickets faster and more effectively than standard LLMs.

In scientific research, Microsoft Discovery used it to grade experiment steps and replan workflows, achieving 46 times higher consistency than LLM evaluations and speeding up adaptive planning by nearly 4x.

Why it matters

Running multi-agent systems on frontier LLMs has always felt like using a rocket engine to back out of a driveway. It works, but it is wildly inefficient.

For developers and system architects, Microsoft-Decision-1 is an immediate win. If your agents waste hundreds of milliseconds evaluating guardrails, routing prompts, or validating JSON fields, dropping this into your pipeline will noticeably lower latency and slice API bills to near zero.

For the broader industry, this launch cements the shift toward model specialization. Big generative models will handle the creative and deep-reasoning heavy lifting, while hyper-specialized, tiny models handle orchestration, guardrails, and control flow.

We think Microsoft's play here is exceptionally sharp. By decoupling decision scoring from open-ended generation, they are tackling the single biggest bottleneck holding back reliable autonomous agents. The only trade-off to watch is whether a 9B model eventually hits a ceiling on nuanced, highly subjective domain rulings, but for standard routing and logic gates, it looks like a no-brainer.