Microsoft-Decision-1

Scored decisions at subscription price.

Microsoft-Decision-1 is the decision model in Microsoft Foundry and on OpenRouter. It scores a fixed set of options instead of writing text. Input $0.042 per million tokens. Output free. No setup fee.

$0.042 / 1M input · output free Post-trained on Qwen3.5-9B Foundry, OpenRouter, Fikra Released 2026-10-09
Why decision models

One-pass choices. No wrangling.

Most application logic doesn't need a generated response. It needs a confident answer from a fixed set of options. Decision models return probabilities per option in a single pass, at a tiny fraction of the cost of an LLM call.

Where it fits

  • Routing: which tier, which model, which team.
  • Content classification: urgency, sentiment, department, priority.
  • Agent gates: continue, stop, escalate, fetch context.
  • Beam scoring: grade candidates before a human review.
  • Policy filtering: do not auto-action; force review below a floor.

Numbers Microsoft published

Benchmarks
83.5% across 36 tests, ~150k questions (internal, unnamed)
Latency
85ms median on Microsoft's tests; Microsoft's claim is "2.5x faster than the runner-up"
Stability
Decision shifts on wording changes: 1.3% average
Price
$0.042 per 1M input tokens. Output free.
Availability
GA on Microsoft Foundry Global Standard, Data Zone Standard; also OpenRouter (provider Azure); also Fikra API.
Example

One input, multiple scored options.

# Microsoft-Decision-1 via Fikra API curl -X POST https://api.fikraapi.co.ke/v1/decisions \ -H "Authorization: Bearer fk_live_..." \ -H "Content-Type: application/json" \ -d '{ "input": {"text": "Customer demands a refund for the July invoice."}, "questions": { "urgent": ["yes", "no"], "team": ["billing", "support", "product", "sales"] }, "model": "microsoft-decision-1" }'

Response shape: one probability per option per question. Use your own floor.

Notes

Microsoft-Decision-1 is on Fikra's pilot roadmap. Ask on [email protected] to be on the first batch.

Find the right threshold. Run 100–200 labeled examples through the model, plot calibration by option, then pick your floor. "Highest confidence" is only the start.
Questions

Common questions.

Is it on Fikra?

Yes. Paid accounts can call it. Pilot accounts as well. Ask on [email protected].

What is it built on?

Post-trained Qwen3.5-9B, Apache 2.0. Microsoft plans to rebase it onto OpenAI and MAI models later.

What's the context limit?

32K tokens. Decisions should fire on short inputs; for very long inputs either chunk the input or use a chat model to extract first.

What is the difference between Gemini Decision 1 and this?

Gemini is a native Google Decisions API with 1.05M context and $0.12 / 1M. Microsoft is $0.042, 32K context. Different price points for different inputs.

Related: Gemini Decisions API guide · All decisions models · Perplexity Decider