Every major cloud platform now offers a purpose-built model for routing, grading, and gating agentic AI workflows, after Microsoft launched Decision-1 on October 9 — completing a category that TypeSafe AI effectively invented 24 days earlier. The entry establishes the decision model not as a niche evaluation tool but as a contested platform layer: a structural chokepoint that determines which queries reach expensive frontier models and which are handled in milliseconds at a fraction of the cost.

Decision-1 is a 9-billion-parameter model post-trained from Alibaba's open-weight Qwen3.5-9B specifically to score decisions rather than generate text. It accepts a context and a fixed set of answer options — yes/no, multiple-choice, or a numeric scale — and returns a calibrated probability for each option in a single forward pass, without producing a word of free-form output. Microsoft claims it is 35 times faster than GPT-6 Sol and more than 100 times faster than GPT-5.6 Luna on evaluation tasks at P50 latency, with the highest accuracy across 36 benchmarks spanning nearly 150,000 held-out questions. All figures are vendor-reported; independent benchmarkers have not yet published a full evaluation, and Microsoft itself recommends teams test on their own data before replacing existing classifiers or routers.

What a Decision Model Actually Does — and Why It's Not Just a Faster LLM

Standard language models generate text token by token. Every forward pass through the network produces one token; a 30-word reply requires 30 or more sequential passes. That architecture is ideal for open-ended reasoning and drafting but wastes computational cycles on tasks where the answer is already constrained — routing a support ticket to the right queue, grading whether an AI-generated response met its rubric, or deciding whether a user request should proceed to a human reviewer. Decision models are built for routing, grading, and gating decisions of exactly this kind.

Decision models handle those constrained tasks through a fundamentally different architectural path. They perform only the prefill phase of transformer inference — processing the full input context in a single parallelizable pass — and then score the probability distribution across the predefined answer options directly, without entering the autoregressive decode phase at all. The result is that a decision requiring what would otherwise be a 30-token explanation completes in one forward pass. At 9 billion parameters, Decision-1's P50 latency reflects both the model size reduction from frontier scale and the elimination of sequential decoding.

The economic consequence follows directly from the architecture. Because no tokens are generated, output billing disappears entirely. Decision-1 is priced at $0.042 per million input tokens on Microsoft Foundry, with output provided at no charge — the same price per input token as TypeSafe's Jev, establishing what appears to be the market-clearing price point for this category. For a workflow making 1,000 decision calls per second at an average of 500 tokens each, the cost works out to roughly $75 per hour — a fraction of what equivalent calls to a frontier model would cost.

Speed Is the Business Case — But Calibration Is the Engineering Requirement

Microsoft frames Decision-1's value primarily around latency. In agentic systems, decisions are rarely isolated: a single workflow may chain 20 or more sequential choices, each gating the next. Adding 100 milliseconds to each of 20 sequential decisions adds two seconds to the overall workflow — a compounding tax that grows with the system's complexity, as the Decision-1 launch announcement details.

But engineers who have worked with the first generation of decision models point to calibration, not raw speed, as the deciding factor in whether the category is safe to deploy. A model that is 96% accurate but overconfident is operationally worse than one that is 94% accurate with honest probabilities, because code acting on overconfident thresholds will fail on cases the model doesn't flag for review. The distinguishing property of both Jev and Decision-1 is that they claim to return calibrated probability scores — meaning a confidence of 0.9 should be right roughly 90% of the time on representative cases, enabling downstream code to set actionable thresholds rather than treating every output as a binary certainty.

Microsoft's robustness testing addresses a related concern: whether the model is responding to semantic content or surface-level features. The company reports perturbation testing each request in eight different ways — paraphrasing option descriptions, reordering choices — and measuring how often the decision flips. Decision-1 changes its answer on just 1.3% of perturbations on average, with zero flips for paraphrased or reordered options.

What Is a Decision Model?

Decision models are not traditional classifiers, which require retraining whenever the set of labels changes. They are not generalist LLMs coerced into returning structured output through prompt engineering — an approach that adds parsing overhead, produces outputs that can fail silently when the model deviates from the expected format, and does not deliver calibrated confidence scores. A decision model accepts typed questions at inference time: the categories to route between can be defined per request, without retraining, as long as they fit within the context window.

The category originated on September 15, 2026, when TypeSafe AI launched Jev — a hosted decision model trained with a method the company calls Reinforcement Learning for Calibrated Decisions, or RLCD. Jev processes a state (a string or structured data) and a set of typed questions in a single parallel evaluation pass, returning probability distributions for each question without sequential token generation. TypeSafe says it attracted more than 150,000 waitlist signups in its first 24 hours and surpassed 1 trillion input tokens per day within the week. Those figures have not been independently verified, and an independent community analysis of user posts about Jev found that real-world speedups had a median of 7 times over the LLM baseline — meaningful, but significantly below the headline 193-times figure circulated in some coverage.

Cloudflare followed on October 1 with Clef and Clef-flash — two open-weight models (27B and 9B respectively) built on the Qwen family, hosted on Workers AI, and released under an Apache 2.0 license, making them freely available for self-hosting. Amazon shipped Strands Decider 2B the same day. Microsoft arrived eight days later, completing a category that did not exist a month earlier.

Inside the Deployments: Xbox, Copilot, and Scientific Discovery

Microsoft says Decision-1 is already running in production across several internal teams, and the early figures illustrate how the architectural trade-off plays out in practice.

Xbox Research used Decision-1 to process more than 10,000 open-ended pieces of player feedback from surveys, Steam, and X, sorting them into researcher-defined themes. The team found Decision-1 competitive in quality with GPT-6 Sol while running over 14 times faster and approximately 200 times less expensive. The Copilot team deployed it for quality control of chat and agentic responses, finding Decision-1 competitive with GPT-5.6 Luna at 100 times the speed.

The most technically significant deployment is Microsoft Discovery, the company's scientific research AI platform. Discovery uses an adaptive replanning loop: an AI agent evaluates an experiment against a rubric, revises its approach based on the rubric grades, and repeats until objectives are met. Swapping Decision-1 into the scoring step yielded 46 times more consistent results than the prior LLM-based scorer at three times the speed — and nearly four times the overall throughput on adaptive replanning cycles. For experiments that might otherwise require days of iterative scoring before converging, that throughput multiplier has direct implications for research timelines.

All internal deployment figures are vendor-reported, derived from Microsoft's own teams using Microsoft's own models as comparison baselines. Independent third-party testing of Decision-1 against the same workloads has not been published.

Who Built This — and Why It Matters That He Did

The announcement came from Achint Srivastava, VP of Software Engineering in the Office of the CTO at Microsoft, whose path to this product traces directly through AI evaluation infrastructure. Srivastava co-founded Pi Labs, an AI evaluation startup acquired by Microsoft in December 2025, whose work focused on the same structured scoring problem that Decision-1 now addresses at scale. His authorship of the announcement reflects a founder's throughline: the acquisition brought both the team and the intellectual framework for building evaluation as infrastructure rather than as an ad hoc layer on top of generative models.

CEO Satya Nadella amplified the launch on X, framing Decision-1 as a tool for "helping people guide and control AI agents through complex environments." That framing aligns with the September 2026 Humanist AI guidelines — a 37-page governance document that places human authority above AI capability and frames AI models as tools rather than autonomous agents. Decision-1 is, in that sense, not just a product bet — it is an architectural expression of Microsoft's stated position that humans should retain control of the decisions agentic systems execute.

Where Decision-1 Sits in a Competitive Landscape Still Taking Shape

No two decision models in the current field are built exactly alike. TypeSafe's Jev is a hosted-only, closed-weight service; Cloudflare's Clef and Clef-flash are open-weight and Apache-licensed. Amazon's Strands Decider 2B removes the language model head entirely, replacing it with a scoring head of just over a million parameters — an approach that makes it smaller and faster but unable to leverage the contextual representations of the language model backbone for nuanced classification tasks. Microsoft post-trained an existing open-weight model (Qwen3.5-9B) rather than building from scratch, and plans to rebase Decision-1 on its own Microsoft AI and OpenAI models in future releases.

On the JevBench Decision Index scores — the 30-plus model benchmark that has become the de facto comparison standard for the category — frontier LLMs still lead on raw accuracy. GPT-6 Luna scores 0.996; decision models in the current generation score 0.83 to 0.93 on JevBench. The accuracy gap is the trade-off: decision models give up some precision in exchange for orders-of-magnitude lower cost and latency, and calibrated confidence scores that allow downstream code to handle the uncertain cases through escalation rather than blind action.

That trade-off becomes sustainable — and the category makes economic sense — when developers design their workflows around it: using decision models for the high-volume, lower-stakes routing and grading calls, and reserving frontier models for the open-ended reasoning steps that require full generative capability. The decision layer, in this architecture, is not a substitute for a frontier model; it is the gate in front of one.

What This Means for Enterprise AI Architecture

The emergence of decision models as a platform category reflects something broader than a product market: it reflects a maturation in how engineering teams think about deploying AI at workflow scale. Rather than routing every decision through the same frontier model used for drafting and reasoning, engineers are reaching for a tiered architecture — heavyweight models for generation and complex reasoning, lightweight decision models for classification, routing, and gating. At $0.042 per million input tokens with output free, the decision layer is economically viable at every node of a workflow rather than rationed to the most consequential ones.

But the more significant implication is architectural control. In a multi-step agentic workflow, the decision model is the entity that determines what reaches the generative model — what question gets passed to GPT, to Claude, to Gemini. Whoever owns that routing layer owns the chokepoint between the user's intent and the frontier model's response. Microsoft's Decision-1, deployed on Azure Foundry, positions Microsoft to be the routing infrastructure for enterprise agentic workflows that also call OpenAI models through the same platform. That is not a neutral infrastructure play — it is a structural claim on the AI enterprise stack that will compound as agentic deployment matures.

"With agents taking action and making an impact in the real world," Srivastava wrote in the Decision-1 launch announcement, "decision models have the potential to help people guide and control those agents through complex environments."

Decision-1 is available on the Microsoft Foundry model catalog and is already available on OpenRouter. Documentation is at Decision-1 documentation on Azure.

Frequently Asked Questions

What is a decision model, and how does it differ from a large language model?

A decision model is an AI model built to score a fixed set of answer options — yes/no, multiple-choice, or a numeric scale — rather than generate free-form text. Where a language model produces its answer one token at a time through a sequential decoding process, a decision model performs a single forward pass over the input and scores all predefined options in parallel. Nothing is generated, so there is no per-token latency and no output billing. The result is a calibrated probability for each option: a confidence score that, when properly trained, should match real-world accuracy — a 90% confidence prediction should be correct roughly nine times out of ten, enabling code to set thresholds for automated action, escalation, or human review.

How does Microsoft Decision-1 compare to TypeSafe Jev and Cloudflare Clef?

All three share the same pricing anchor for input tokens ($0.042/M for both Decision-1 and Jev; Clef charges more at $0.24/M for the 27B model, $0.09/M for Clef-flash). The differences are in architecture and availability: Jev is a hosted, closed-weight service; Clef and Clef-flash are open-weight models under Apache 2.0, available for self-hosting. Decision-1 is hosted on Microsoft Foundry (Azure). On the JevBench Decision Index, current-generation decision models score 0.83 to 0.93 accuracy; frontier LLMs score higher (GPT-6 Luna: 0.996), but at orders-of-magnitude greater cost and latency. All benchmark figures in the category are currently vendor-reported; independent third-party evaluations are not yet available.

Is the decision model's "control layer" role a concern for enterprise teams already committed to specific frontier AI vendors?

It should be a factor in vendor evaluation. In a tiered agentic workflow, the decision model determines which queries reach the generative model — making the routing layer a structural dependency. An enterprise deploying Decision-1 on Azure Foundry to gate calls to OpenAI models is concentrating both the routing control and the inference infrastructure on Microsoft's platform. Teams that want to maintain flexibility across frontier providers should evaluate whether their decision model can route to multiple generative backends, and whether their workflow orchestration layer can substitute a different decision model if pricing or policy changes.

Can a decision model handle tasks that require nuanced judgment, or only binary routing?

Decision models are suited to tasks with a defined, enumerable answer space — classification, routing, grading, prioritization, and rubric-based evaluation. They perform poorly on tasks requiring open-ended reasoning, arithmetic, or date comparison; known weaknesses in current-generation models include unreliable counting and accuracy loss on large, noisy input states. For tasks at the edge of those limits, the recommended architecture is threshold-based escalation: let the decision model handle cases above a confidence threshold automatically, and route low-confidence cases to a larger model or a human reviewer. The decision model's calibrated probability output is what makes that escalation logic actionable.