TL;DR — Key Takeaways
- Fast decisions are becoming a separate AI workload. Microsoft’s Decision-1 model handles classification, routing, prioritization and verification through structured probability scores rather than generating free-form text, helping agents make small decisions faster.
- Microsoft is challenging TypeSafe AI on performance and distribution. Decision-1 competes directly with Jev at the same advertised input-token price of $0.042 per million tokens. Microsoft claims superior benchmark accuracy and latency, although independent validation remains limited.
- Speed cannot replace reliable verification. Decision models can inherit mistakes from the larger models used to train or evaluate them. Enterprises should test calibration against their own labeled data, establish confidence thresholds and retain human oversight for consequential decisions.
Microsoft thinks AI agents need a second kind of model. Not one that writes or reasons. One that makes quick calls.
The company has released Microsoft-Decision-1, a decision-scoring model available in Microsoft Foundry and through OpenRouter. It’s built for routing, classification, prioritization, verification, and workflow control. And it puts Microsoft in direct competition with Jev, the decision model TypeSafe AI shipped less than a month ago.
What a Decision Model Does
LLMs generate text. Decision models don’t. You give a decision model a fixed set of answers, and it returns a calibrated probability for each one. Software can act on that output right away. There’s no free-form text to parse.
Decision-1 handles yes/no questions, multiple choice, and ratings. It can also grade AI responses and agent actions against a rubric. All of it runs through one structured API call.
Microsoft built the model by post-training Qwen3.5-9B, an open-weight model with 9 billion parameters, for single-pass scoring. The company plans to rebase it on its own MAI models and on OpenAI models.
Why Speed Matters
Latency adds up in agent workflows. Microsoft gives a simple example. Add 100 milliseconds to each of 20 sequential decisions, and you’ve added two seconds to the job.
Agents make many small calls. Retry or stop? Which tool next? Does this output pass? Sending each of those to a frontier model is slow and expensive.
Microsoft says Decision-1 posted the highest accuracy across 36 benchmarks, covering nearly 150,000 questions kept out of training. It also says the model ran 2.5 times faster than the runner-up, H2O-Lightning-4B, and about 35 times faster than GPT-6 Sol at median latency.
These are Microsoft’s own numbers. Independent testing hasn’t caught up yet.
Two other claims stand out. First, robustness. Microsoft rewrote each test request eight different ways and measured how often the answer changed. On average, it flipped 1.3% of the time, and never when options were reordered or reworded. That matters because position bias is a known weakness in LLM-based classifiers.
Second, calibration. The probability is part of the product. If the model says 90%, it should be right about nine times out of 10. Applications use that number to decide whether to act, defer, or send work to a human.
The Jev Factor
TypeSafe released Jev in early access on September 15, 2026, and describes it as the first public System One model. The name borrows from Kahneman’s dual-process theory: Fast, bounded judgment rather than slow, deliberate reasoning. The industry settled on the simpler label, “decision model.”
Jev picked up users quickly. Vercel replaced a ChatGPT-based safety classifier with Jev and reported results five to 18 times faster, with better accuracy.
Microsoft was clearly watching. Its announcement leans on JevBench, an open leaderboard, and an editor’s note says the post was updated to add Jev results for accuracy and calibration. Then there’s price. Microsoft charges $0.042 per million input tokens, and output tokens are free. That’s the same price TypeSafe lists for Jev.
So Microsoft isn’t trying to undercut TypeSafe. It’s competing on benchmarks and distribution. Plenty of enterprises already run on Azure and Foundry. For them, Decision-1 is one click away in the catalog.
How Microsoft Uses It Today
Microsoft shared a few internal tests. Xbox Research sorted more than 10,000 pieces of player feedback into themes. Quality matched GPT-6 Sol, at 14 times the speed and a fraction of the cost. The Copilot team used it to grade chat and agent responses, reporting quality close to GPT-5.6 Luna at 100 times the speed.
Microsoft Discovery, the company’s science platform, uses rubric grades to replan experiments. Microsoft says Decision-1 scored far more consistently than the LLM it replaced and sped up the replanning loop nearly fourfold. On-call engineers also use it to pull relevant knowledge during live incidents.
The Caveats
Decision models are useful. They aren’t magic.
Amplitude tested Jev and found the expensive LLMs were more accurate than the decision model in its benchmark. It also cited a study where the decision model agreed with the frontier model 6 to 8 points more often than it matched the ground-truth labels.
That gap is the real risk.
“A decision model tuned to match a frontier model can inherit that model’s mistakes and repeat them at a fraction of the latency,” said Mitch Ashley, vice president and practice lead for CIO & Technology Buyers and Software Lifecycle Engineering at The Futurum Group. “Faster grading makes verification debt worse when the grader and the work it grades share the same blind spots.”
Microsoft adds its own guardrail. The Foundry catalog warns against using Decision-1 as the only automated decision-maker for consequential matters like employment, credit, housing, healthcare, or legal rights. And it’s text-only for now.
What Teams Should Do
One simple test applies: If you can define the valid answers before the model runs, a decision model may be worth a look.
Start with one high-volume decision. Routing, safety screening, and output grading are good candidates. Then test it on your own data before you trust it.
“Teams need their own labeled ground truth and a pinned model version before they trust a score,” Ashley said. “Microsoft and TypeSafe charge the same price, so calibration on customer data, outside vendor benchmarks, will decide which one enterprises standardize on.”
From there, set confidence thresholds for when to escalate to a person or a larger model. And keep checking. A score that holds up in week one won’t automatically hold up in month six.
Agents need a fast layer for small calls. Two vendors now offer one at the same price. The next few months will show whose numbers hold up outside the lab.
