Decision-1: Microsoft's $0.042/M Token Classifier

Microsoft released Decision-1 on October 10, 2026. The model promises calibrated probabilities in one pass at a price of $0.042 per million tokens, according to Grok AI News. The release signals a deliberate shift toward task-specialized decision models over general-purpose generation.

Model: Decision-1 | Base: Qwen3.5-9B
Key feature: Calibrated probabilities in one pass | Price: $0.042 per million tokens

What It Is / How It Works
Decision-1 is a classifier built on the Qwen3.5-9B foundation designed for routing and verification tasks rather than free-form text generation. It outputs calibrated probability distributions for decision labels in a single inference pass, enabling direct, probabilistic routing decisions without invoking a separate generator. The one-pass mechanism is positioned to reduce latency and cost compared with multi-step or generation-heavy pipelines. In practice, teams can deploy Decision-1 to decide whether an input should be approved, routed to a human reviewer, or escalated to a different workflow, all with a single probabilistic output vector. The approach aligns with a broader shift toward specialized decision models that aim to outperform full LLMs on constrained decision tasks while preserving reliability through calibrated scores. The core value proposition is simple: fewer tokens, more actionable certainty, and lower spend.

Benchmarks / Specs / Numbers
Decision-1 tops 36 benchmarks across decision-oriented tasks, signaling strong performance on routing and verification workloads relative to similarly sized classifiers. Quick specs (from the source material): the model is based on Qwen3.5-9B, outputs calibrated probabilities in one pass, and costs $0.042 per million tokens. These data points position Decision-1 as a cost-conscious alternative to full LLMs when the task is classification, labeling, or routing rather than content generation. For readers measuring practicality, the 36-benchmark success claim is a meaningful proxy for reliability across a range of decision tasks—from document routing to compliance screening. For context on calibrated outputs versus generation, see background work on probability calibration in classifiers and neural nets. Background reading: Calibration in ML and foundational research like On Calibration of Neural Networks.

How to Try It
If Microsoft opens access to Decision-1 via API or on-premise tooling, here’s a pragmatic path to evaluation:

• Step 1: Verify availability through official Microsoft channels or the cited Grok AI News article, then request access or sign up for the API/SDK as appropriate.
• Step 2: Obtain an API key or local deployment package and install the client library recommended by the official docs.
• Step 3: Prepare a decision-task dataset (inputs with target labels for routing or verification). Pass each input to Decision-1 and collect the probabilistic outputs.
• Step 4: Assess calibration and decision quality. Use reliability diagrams and Brier scores to quantify probability calibration, and compare decision accuracy against a baseline classifier or a small LLM-based approach. Helpful background on calibration techniques: Calibration in ML and Calibrated Neural Networks.
• Step 5: Benchmark cost against your current stack. At $0.042 per million tokens, run a cost-per-task calculation to understand break-even points versus full LLM generation.
• Step 6: Compare to alternatives. If you need generation or richer reasoning, consider OpenAI’s GPT-4 or Anthropic’s Claude as generative baselines, noting that both typically incur higher per-token costs than a dedicated decision classifier (OpenAI: GPT-4, Anthropic: Claude).
• Step 7: Iterate with domain-specific calibration. If your domain includes regulatory or safety concerns, incorporate additional calibration or thresholding to meet risk criteria.

Pros and Cons

• Pros:
  • Cost efficiency: designed as a low-cost alternative to full LLMs for decision tasks (price point cited as $0.042 per million tokens).
  • Latency and simplicity: one-pass calibrated probabilities streamline routing and verification without generating long text.
  • Calibrated outputs: probabilistic scores enable principled thresholding and risk-aware decisions.
• Cons:
  • Task scope: optimized for decision tasks; not a drop-in replacement for generation or complex reasoning workloads.
  • Domain adaptability: effectiveness depends on domain alignment with the 36 benchmark set; cross-domain performance can vary.
  • Ecosystem openness: practical access and integration details are contingent on Microsoft’s API/docs rollout, which affects ease of adoption.

Alternatives and Comparisons
To gauge when Decision-1 makes sense, compare it to two common alternatives: a full generative LLM and a different decision-focused approach.
| Feature | Decision-1 (Qwen3.5-9B classifier) | GPT-4 (OpenAI) | Claude (Anthropic) |
|---------|-----------------------------------|-----------------|-------------------|
| Primary use | Decision routing and verification with calibrated probabilities | General generation and reasoning | General generation and reasoning |
| Output type | Probabilities for decision labels | Generated text and reasoning | Generated text and reasoning |
| Cost model | $0.042 per million tokens (classifier usage) | Per-token generation cost (higher) | Per-token generation cost (higher) |
| Latency | Typically lower for decision tasks (one-pass) | Higher due to generation | Higher due to generation |
| Calibration focus | Yes (calibrated probabilities in one pass) | No inherent calibration emphasis for decisions | No inherent calibration emphasis for decisions |
Who should use it vs. alternatives: if your objective is fast, scalable decision routing with calibrated confidences and you can tolerate a fixed decision label set, Decision-1 is attractive. If you need richer language generation, creative tasks, or extensive multi-turn reasoning, a full LLM like GPT-4 or Claude may be more suitable, albeit at higher cost and latency. The decision hinges on whether your workflow benefits from probabilistic routing signals or requires fluent natural language generation. For readers seeking broader context on probabilistic calibration in models and industry adoption, see arXiv: Calibration of Neural Networks and Wikipedia on Probability Calibration.

Who Should Use This

• Enterprises with high-volume routing or verification tasks and strict cost controls.
• Teams needing deterministic, probabilistic outputs to gate downstream workflows (e.g., fraud screening, compliance triage, document routing).
• Projects where latency must be minimized and the decision space is well-defined and finite.
• Organizations evaluating a shift from costly generation-heavy pipelines to specialized decision models.
• Skip if your tasks require deep language generation, complex reasoning with nuanced narratives, or multi-turn interactions beyond simple decision labels.

Bottom Line / Verdict
Decision-1 positions Microsoft at the intersection of cost-conscious efficiency and reliable decision-support by delivering calibrated probabilities in a single pass on a 9B-scale model. Its 36-benchmark performance claims and the low $0.042 per million tokens price make it an appealing option for scalable routing and verification workloads where generation is not the primary goal. Adoption success will hinge on real-world access, domain adaptation, and how well calibrated outputs translate into reliable business decisions compared with both traditional classifiers and full LLMs. In short, Decision-1 is a practical unlock for decision-heavy pipelines—worth evaluating when cost and speed are priorities, but not a wholesale replacement for generative AI in tasks requiring fluent language and broad reasoning. As the market continues to separate decision models from generation models, Decision-1 could become a standard building block for enterprise decision workflows.

CLOSING
As specialized decision models mature, expect more cost-conscious options to outperform general-purpose generation in constrained tasks. Decision-1’s pricing and one-pass calibration make it a notable datapoint in this ongoing shift toward task-focused AI tooling.