Microsoft Releases Decision-1, a Small Model for Classification and Agent Routing

Microsoft on October 9, 2026 released Microsoft-Decision-1, a small model that returns a calibrated probability for each of a fixed set of answer options instead of writing text. It was the most accurate and the fastest model in Microsoft's own 36-benchmark comparison, the company said, and it costs $0.042 per million input tokens, with output tokens free.

Decision-1 is the latest of several decision models released since mid-September. TypeSafe AI introduced Jev, the model that started the category, in mid-September (The Decoder), Cloudflare followed with Clef on October 2, and OpenAI opened its Decisions API to all developers on October 6. Like other recent small models, these compete on speed and price rather than broad ability, and TypeSafe announced on October 9 that it had raised $870 million at a $7.5 billion valuation in a round led by Andreessen Horowitz.

What did Microsoft release?

Microsoft-Decision-1 is a decision-scoring model that Microsoft post-trained from Alibaba's open weight Qwen3.5-9B, announced on October 9, 2026 by Achint Srivastava, VP of Software Engineering in Microsoft's Office of the CTO. Given a situation and a closed question, it scores every allowed answer in a single pass.

A decision model is a model that picks from answers a developer defines in advance and attaches a probability to each, so software can act, defer or ask for review based on that number. Decision-1 handles yes or no questions, a choice among named categories and ratings on an ordered scale, and it can grade AI responses and proposed agent actions against a rubric, according to Microsoft's announcement.

The Foundry catalog entry lists the model as generally available, text only and limited to inputs of 32,768 tokens. Microsoft says it will "soon rebase" Decision-1 on other models, including its own MAI models and OpenAI's, and OpenRouter's listing says the weights are updated continually while the API shape stays the same.

How does Decision-1 compare with Jev and OpenAI's model?

In Microsoft's own comparison, Decision-1 has the highest average accuracy of the 8 models with an accuracy score and the lowest median latency of the 8 with a latency figure, but ranks third of 7 on calibration. Calibration measures whether a model's confidence matches how often it is right.

The accuracy lead over Jev works out to 1.2 percentage points. The latency figures for rivals come from JevBench v1.6.1, checked on October 7, while Decision-1 was measured through Foundry, according to the chart's footnotes. Microsoft added the Jev row after first publishing the post, its editor's note says. Cloudflare's Clef models were not part of the comparison (The Decoder).

Speed claims in this category do not line up across vendors. Microsoft's chart puts Jev's median latency at 240 ms, while Cloudflare reported just over 524 ms for Jev when it launched Clef (The Decoder). Each vendor picks its own test setup, so the numbers cannot be compared directly.

Microsoft also reports that Decision-1 changed its answer on 1.3% of reworded or reordered requests on average, and that it was tested on 5,250 safety requests across 11 benchmarks.

How much does Decision-1 cost, and where is it available?

Decision-1 costs $0.042 per million input tokens, with output tokens free, as of October 2026, Microsoft says. That is the same list price TypeSafe gave for Jev at launch (The Decoder) and less than half of the $0.10 per million input tokens OpenAI charges for its Decisions API with GPT-6 Luna, which also has no output charge.

The model runs in Microsoft Foundry and through OpenRouter, where Azure is the only provider. Vercel added it to its AI Gateway on October 9, callable through OpenAI compatible and TypeSafe compatible APIs as well as its own AI SDK.

In a cost chart, Microsoft estimates that classifying 1 million texts would cost about $11 with Decision-1 and about $2,434 with GPT-6 Sol, its large reference model.

How has Microsoft tested it internally?

Microsoft says several of its own teams have tested Decision-1, and the results it reports are its own, not independent checks. Xbox Research sorted more than 10,000 pieces of feedback and reviews into fixed themes and found Decision-1 competitive on quality with GPT-6 Sol while running over 14 times faster at a cost 200 times lower, the company says.

The Copilot team found it "competitive with GPT5.6 Luna and 100 times faster," according to Microsoft. That comparison uses an older Luna model than GPT-6 Luna, which appears in Microsoft's own benchmark chart. In Microsoft Discovery, where an agent grades and revises its own experiments, Decision-1 scored 46 times more consistently than an LLM-based grader at three times the speed, Microsoft says.

What we don't know yet

• How Decision-1 performs in independent tests, and how it compares with Cloudflare's Clef.
• Which Azure regions offer the DataZoneStandard deployment that keeps processing inside one data zone, including in Europe.
• When Microsoft will rebase the model on MAI or OpenAI models, and whether its scores and price will change then.
• How results shift over time, since the weights are updated continually.