Microsoft's New AI Doesn't Write A Word. It Just Decides.

Today, we're diving into:

Microsoft's New AI Doesn't Write A Word. It Just Decides.

Most enterprise AI conversations are still about models that write: emails, code, summaries, slides. Microsoft just shipped one that refuses to.

On October 9, Microsoft released Microsoft-Decision-1, a model built for routing, classification, verification and agent control. Instead of generating text, it "returns a calibrated probability for each fixed answer option."

Your Agents Need Rules They Can Check Against.

Find out whether your data, definitions and policies are ready to become the evidence AI decisions rely on.

A Model That Picks From A Menu

You give it a situation, a question and a fixed set of options. It scores every option in a single call and returns the probabilities as JSON.

No essay. No rationale. Just a ranked answer and how confident the model is in it.

Microsoft lists the jobs it is built for: yes/no and multiple-choice questions, rubric scoring, grading AI responses and proposed agent actions, and checking whether an answer is grounded in the evidence supplied.

It also supports explicit options such as "cannot tell," so the model can abstain instead of guessing. That small design choice matters more than any benchmark.

The Numbers Microsoft Is Claiming

Across 36 benchmarks and more than 147,000 questions, Microsoft reports 83.5% average accuracy, ahead of OpenAI's GPT-6 Luna Decisions at 79.4%. It puts median latency at 85 milliseconds, and pricing at $0.042 per million input tokens, with output tokens free.

Internally, Microsoft says Xbox Research used it to sort more than 10,000 feedback items 14 times faster and 200 times cheaper than GPT-6 Sol.

Why This Matters More Than Another Chatbot

Here is the shift most leaders will miss. As agents take on real work, the expensive question is no longer "what should the agent write?" It is "should the agent be allowed to do this?"

Today, many teams answer that question by asking a large frontier model to review another model's output. That works. It is also slow, costly and inconsistent, because a model built to write prose was never designed to be a judge.

This fundamentally changes how you should architect agent control. Generation and judgment are becoming separate layers, with separate models, separate costs and separate owners.

The Wrong Reaction, And The Better One

The wrong reaction is to bolt a decision model onto every workflow and call it governance. A judge is only as good as the options you give it and the evidence it reads.

The better alternative is to treat decision models like a policy engine. You define the choices, set the confidence thresholds and decide what happens when a score lands in the gray zone. Microsoft's own model card is blunt about the limits: it is "Not for sole automated decisions on credit, employment, housing, healthcare or legal rights."

The Real Bottleneck Is The Evidence

A groundedness check asks one simple question: does this answer match the evidence? That only works if the evidence itself is right.

Picture an agent that proposes a customer refund, and a decision model that checks it against your returns policy. If three systems hold three versions of that policy, the model will confidently approve the wrong thing.

Reliable AI judgment starts with trusted enterprise context. DataManagement.AI helps organizations create that foundation by unifying metadata, lineage, governance, business glossaries and knowledge assets into a single intelligence layer, so every decision model checks agent actions against one approved version of the truth.

Nvidia Is Circling The Open Model You Planned To Run

Open-weight models were supposed to be your escape from vendor dependence. This weekend, the most powerful vendor in AI moved closer to owning one.

On October 10, the Financial Times reported that Nvidia is in talks to acquire Reflection AI or deepen its investment in the US open-weight startup. Reflection was last valued at $25 billion, and deal terms could not be established.

Why Reflection Matters

Days earlier, Reflection unveiled Beam, a 501-billion-parameter open-weight model with 23 billion parameters active at a time and up to 1 million tokens of context. Weights, a technical report and developer tooling are due later in October under an Apache 2.0 license.

Reflection says Beam matches China's GLM-5.2 on advanced reasoning while using 3 to 4 times less inference compute. That claim is the company's own and has not been independently verified yet.

The Twist In The Deal

Nvidia has already invested $800 million in Reflection and supplied the GPUs for Beam's training run, according to RuntimeWire. A deeper deal would put the world's dominant AI chip supplier on both sides of the open-weight market: the hardware you run on and the model you run.

The FT also reported the deal could be structured as an acquihire, a format that can help a buyer avoid antitrust scrutiny.

What This Means For Your Open-Weight Strategy

Open weights do not vanish when a company changes hands. An Apache 2.0 release stays usable. But the next version, the tooling, the optimizations and the support roadmap all follow whoever funds the team.

That matters because open models now carry real enterprise load. Constellation Research, citing Vercel AI Gateway data, found open-weight models were 29% of token volume in June but under 4% of spend, and reached 62% of token volume by August.

Open models are where the volume is going precisely because they are cheap. If their roadmaps start optimizing for one vendor's hardware, that price advantage could quietly narrow.

The Better Alternative

The wrong reaction is to abandon open weights. The better alternative is to treat them as one tier in a portfolio, chosen workload by workload, with a tested path to move.

Know Your Inference (KYI) is becoming a core enterprise capability. By understanding how every inference affects cost, latency, accuracy and governance, you can see which workloads truly benefit from open models, and switch providers on evidence rather than headlines.

OpenAI's Revenue Just Came In $20 Billion Lighter

The AI boom runs on one closely watched number. Last week, that number shrank.

The Financial Times reported that OpenAI told investors its annualized revenue was close to $50 billion at the end of September, about $20 billion below the roughly $70 billion figure that media reports had circulated weeks earlier. A source told Reuters the gap "mainly arose from an attempt to produce a direct comparison" with Anthropic's figures.

Same Metric, Two Different Rulebooks

According to Bloomberg, the two labs calculate this closely watched revenue figure differently. Anthropic books gross sales through cloud partners. OpenAI records only its net share.

Reuters adds that Anthropic pays cloud partners about 16% of each dollar earned through them, and that cloud-channel sales made up half of Anthropic's revenue last year. In the second quarter, Anthropic reported $11.5 billion against OpenAI's $6.7 billion, the first quarter it came out ahead.

The Market Noticed

AI-linked stocks fell after the report, with Nvidia down 2.9% and Oracle down nearly 6%, according to AFP. Per the FT, OpenAI is negotiating new private funding at a valuation of roughly $1.4 trillion, and has postponed its expected listing.

Why This Lands On Your Desk

You are probably not an OpenAI investor. But you are likely signing multi-year AI commitments based on assumptions about which vendors will grow, keep cutting prices and stay independent.

Annualized run rate is a fragile number, because it often takes one month of sales and multiplies it by 12. If a $20 billion swing can hide inside the metric investors watch most closely, the growth stories in your vendor pitch decks deserve the same scrutiny.

The Same Trap Lives Inside Your AI Program

Picture your next quarterly AI review. One team reports hours saved. Another reports tickets deflected. A third annualizes a two-week pilot.

Every number looks good. None of them add up to the same thing. The real bottleneck is not the metric. It is the missing shared definition of what counts.

Your opinion matters!

Hope you loved reading our piece of newsletter as much as we had fun writing it.

Share your experience and feedback with us below 'cause we take your critique very critically.

Thank you for reading

-Shen & Towards AGI team