NewsStocksMicrosoft-Decision-1.1 Scores 68 on Artificial Analysis Decision Index

Microsoft-Decision-1.1 Scores 68 on Artificial Analysis Decision Index

Author: CryptoBriefing·

Key Takeaways

  • •Microsoft-Decision-1.1 launched on November 17, 2026, and scored 68 on the Artificial Analysis Decision Index v2.1, an improvement over its predecessor's score of 62.
  • •The new model completes tasks in 123 milliseconds compared with 274 milliseconds for Decision-1, more than doubling speed while keeping the price flat at $0.042 per million input tokens.
  • •Decision-1.1 is a post-trained version of Qwen3.5-9B with 9 billion parameters and a context window of roughly 32,000 to 33,000 tokens, tuned for single-pass probability scoring across yes/no, multiple-choice, and rubric-based formats.
  • •The model ranks eighth among decision models on the independently run v2.1 index, placing it ahead of Jev 1.14, Google's Gemini Decision 1, and OpenAI's GPT-6 Luna, though seven models still score higher.
  • •Microsoft reports that more than 4,000 customers have used the model since its October 2026 public preview and claims top average accuracy across 102 blind benchmarks, though that accuracy figure comes from Microsoft rather than an independent assessor.
Microsoft-Decision-1.1 Scores 68 on Artificial Analysis Decision Index

Microsoft has released Microsoft-Decision-1.1, a specialized AI model built to make business decisions. The model scored 68 on the Artificial Analysis Decision Index v2.1, runs more than twice as fast as the version before it, and keeps its price unchanged.

The first version only entered public preview in October 2026, which means Microsoft shipped a faster, higher-scoring sequel within a matter of weeks.

What changed between versions

The new model launched on November 17, 2026. Its benchmark score climbed to 68, up from the 62 that its predecessor, Microsoft-Decision-1, posted on the same index.

Speed is the bigger headline. Decision-1.1 completes a task in 123 milliseconds, compared with 274 milliseconds for the earlier model, making it more than twice as fast as Decision-1.

Pricing did not move. The model still costs $0.042 per million input tokens, the same rate as Decision-1, so customers effectively get more throughput for the same spend.

How the model actually works

Under the hood, Decision-1.1 is a post-trained version of Qwen3.5-9B, a relatively compact base model. It carries 9 billion parameters and handles a context window of roughly 32,000 to 33,000 tokens.

It is tuned for single-pass probability scoring: give it a question, and it returns a calibrated estimate of how likely each answer is. That design fits the economics of high-volume scoring workloads, where per-task latency and token cost determine whether a model can be run at scale in production.

Microsoft built it to handle several formats. These include yes/no answers, multiple-choice questions, and rubric-based assessments.

Microsoft says it has been using the model internally. The company also reports that more than 4,000 customers have used it since the public preview began. Those adoption figures come from Microsoft.

Where it sits on the leaderboard

Artificial Analysis, which runs independent benchmark assessments, ranks Decision-1.1 eighth among decision models on its v2.1 index. That places it ahead of several recognizable names, including Jev 1.14, Google's Gemini Decision 1, and OpenAI's GPT-6 Luna. Seven models remain ahead of Decision-1.1 on the same index. Because Artificial Analysis runs its assessments independently of the vendors being ranked, the index gives buyers a neutral reference point for comparing specialized models across providers.

Microsoft has also made its own accuracy claims. The company says the model leads on average accuracy across 102 blind benchmarks, with more than 385,000 questions evaluated. That figure comes from Microsoft rather than from Artificial Analysis.

What this means

Microsoft held its token price flat while cutting task time from 274 milliseconds to 123 milliseconds, so customers effectively receive greater throughput for the same spend, roughly doubling the number of tasks completed per unit of time at no additional cost. Because Decision-1.1 is built on a compact 9-billion-parameter base, its placement above entries from Google and OpenAI on the Artificial Analysis index is a notable result.

Still, seven models rank ahead of Decision-1.1 on the v2.1 index, so Microsoft is competitive rather than dominant on that particular scoreboard. Microsoft's broader accuracy claim across 102 benchmarks has also not been framed as an independent result, and buyers evaluating the model will want to test it against their own data before trusting it with decisions that carry real consequences. Two developments are worth watching from here: whether subsequent releases narrow the gap to the seven models above it on the v2.1 index, and whether independent results emerge to corroborate the company's 102-benchmark accuracy claim.