Microsoft builds its decision model on Alibaba’s Qwen, not OpenAI
Microsoft launched Decision-1 in Foundry, a decision model post-trained on Alibaba’s Qwen3.5-9B rather than any OpenAI model, priced at $0.042 per million input tokens. Validate its calibration on your own data before trusting it with sensitive decisions.
9 October 2026. Satya Nadella announces Microsoft-Decision-1, a “fast decision model” already being tested across Microsoft. 9 October 2026. The model is post-trained on Alibaba’s Qwen3.5-9B, not on anything from its partner OpenAI, three days after OpenAI opened its Decisions API to every developer in public beta. 9 October 2026. TypeSafe, the startup that started the category, raises $870 million in a Series A at a $7.5 billion valuation led by a16z. Why it matters: “decision models” — small, fast, calibrated to choose rather than generate — are becoming a commodity layer, and Microsoft picked open Chinese weights as its foundation over its own ally’s model.
What a decision model actually is
A decision model does not generate prose. It answers a closed question — “is this alert critical?”, “does this message violate policy?”, “which action should the agent take next?” — and returns a decision, often with a calibrated probability. The idea, popularised by the “System 1 / System 2” current, is to reserve frontier models for generation and planning, and hand the frequent judgements to a small specialist model that is cheaper and several orders of magnitude faster.
The economics are mechanical. An agent can emit dozens of decisions for a single generated answer; if each decision costs a fraction of a cent and runs in milliseconds, the bill and the latency of the whole system collapse, without degrading quality on the tasks where specialisation is enough. Decision-1 sits exactly in this slot, on a 9-billion-parameter base where a frontier model counts hundreds of billions.
A category three weeks old and already crowded
The decision model category is young, and its founder is neither Microsoft nor OpenAI. It is TypeSafe, whose model Jev defined the use case: a small specialist model that answers structured questions — route a request, grade a response, pick an agent’s next action — faster and cheaper than a large general-purpose model. Microsoft arrives roughly three and a half weeks behind TypeSafe.
In between, everyone piled in. OpenAI opened its Decisions API in public beta, Upstage, Perplexity, Cloudflare and AWS each shipped their own decision model, and AWS, Upstage and Ollama adopted TypeSafe’s System One API as the category’s common interface — a standardisation we covered in our analysis of the Clef and AWS decision models. Decision-1 therefore lands in an already-crowded market, where differentiation comes down to price, base model and reliability.
Why Microsoft skipped OpenAI
The base-model choice is the most telling fact of the announcement. Decision-1 is post-trained on Qwen3.5-9B, Alibaba’s open-weight model, even though Microsoft holds one of the industry’s most strategic partnerships with OpenAI — to the point of having already said it plans to rebase Decision-1 on its own MAI models and then on OpenAI’s. The signal is clear: for such a specialised task, Microsoft did not wait for its partner and took the most accessible open base available.
The Qwen choice is not an outlier. Cloudflare built Clef-flash on the same Qwen3.5-9B, and Jared Palmer, VP of engineering at Cognition, says he spent about $95 of H100 GPU time on Modal porting his Kev models to Qwen3.5. An open base, a negligible adaptation cost, a permissive licence: decision models are becoming the first AI segment where the foundation is a commodity you can swap in an evening.
The choice resonates all the more because the Microsoft/OpenAI relationship is under constant scrutiny. Microsoft has invested tens of billions in OpenAI and hosts its models on Azure; seeing Foundry launch a model built on Qwen, with a rebasing plan that puts MAI before OpenAI, draws a landscape where Microsoft no longer depends on a single model supplier — even its closest partner.
Microsoft’s first customer is itself
Microsoft’s rollout is itself a case study. Xbox Research used Decision-1 to sort more than 10,000 pieces of player feedback, the Copilot team graded chat and agent responses with it, on-call engineers pulled context during live incidents, and Microsoft Discovery scored experiments before an agent replanned. Microsoft’s own numbers claim it is more than 14× faster than GPT-6 Sol at a fraction of the cost for Xbox, and 46× more consistent in Discovery. When a company this size validates a model across four teams before shipping it, the model is not a science experiment — it is infrastructure being hardened.
Jev’s price becomes the reference rate
Decision-1 is priced at $0.042 per million input tokens with free output — exactly what Jev charges. That is no coincidence: TypeSafe’s price has become the category’s de facto benchmark. Jared Palmer lists Kev-4B on OpenRouter at $0.042, Perplexity charges $0.02, and OpenAI sits at more than double Jev’s rate.
At those prices, the margin on an isolated decision call is close to zero. Microsoft’s strategy lies elsewhere: keeping agent traffic — including the generative calls around each decision — flowing through Foundry, and using the OpenRouter listing to pull in developers who do not use Azure. The decision model becomes a loss-leader to sell the rest of the stack.
Read the pricing against the whole category and a sharper picture emerges. When Cognition’s Jared Palmer lists Kev-4B at the same $0.042, and Perplexity undercuts everyone at $0.02, the routing call itself is approaching a zero-margin commodity. The money, for Microsoft, is not in the decision — it is in the generative workload the decision steers toward Foundry.
What Decision-1 does not do yet
Decision-1 trails on two fronts. First, modality: its Foundry listing caps input at 32,768 tokens of text and returns JSON, with no image support — unlike OpenAI’s Decisions API and Clef, which ship a vision encoder. Second, weights: where Cloudflare released Clef under Apache 2.0, Microsoft has announced no open weights. Third, interoperability is still murky: the Foundry sample code calls a /systemone endpoint, but Microsoft has not confirmed full compatibility with the System One API its rivals adopted.
The contrast is instructive. Microsoft builds on an open base but does not open its result — an asymmetry teams that value portability should keep in mind.
The open question of calibration
The last point is the most delicate, because it is about trust. Microsoft says Decision-1’s probabilities are calibrated — a 90% prediction should be right about nine times out of ten on representative cases. But research on Jev shows how far a confident score can drift when the input is written to mislead.
In the JevOut preprint, USC researcher Zixiang Xu and co-authors flipped 312 of Jev’s 508 initially-correct decisions with short, natural-sounding context additions, and Jev assigned at least 70% probability to the wrong answer in 229 cases. Microsoft says it tested Decision-1 with eight perturbation types (reordered options, paraphrased descriptions), which shifted only 1.3% of answers on average — but JevOut did not test Decision-1, and Microsoft itself advises customers to validate calibration on their own data. A confidence score is only usable once it has been stress-tested against your distributions.
Verdict
If you route agents or grade responses at scale and want to stay in the Azure ecosystem, Decision-1 is a serious candidate: Jev-matching pricing, Foundry integration, and claimed internal wins of 14× the speed of GPT-6 Sol at a fraction of the cost for Xbox. If portability or multimodality matters, look at Clef (open, with vision) or TypeSafe’s System One API before committing. In every case, recalibrate on your own data under adversarial pressure before letting a decision model arbitrate a decision that costs money or reputation.