On Friday, Microsoft rolled out Microsoft-Decision-1 through Foundry, arriving just three days after OpenAI made its Decisions API available to all developers in public beta. Rather than relying on OpenAI's technology, Microsoft performed post-training on Alibaba's Qwen3.5-9B, though the company indicated plans to rebase the model onto its own MAI systems and OpenAI offerings in the future.
The pricing mirrors TypeSafe's Jev exactly at $0.042 per million input tokens with no output charges. The launch coincided with TypeSafe's announcement of an $870 million Series A funding round led by a16z, valuing the startup at $7.5 billion. Microsoft trails the category pioneer by roughly three and a half weeks; in that span, OpenAI, Upstage, Perplexity, Cloudflare, and AWS have each introduced their own decision models.
TypeSafe reports that nearly 30% of Fortune 500 companies have experimented with Jev since its launch, though the firm has withheld their identities. TypeSafe CEO Diogo Almeida stated on X that "29.4% of the Fortune 500 showed up" within three weeks of the product's release. These same enterprises represent Microsoft's core Azure customer base, making TypeSafe's rapid adoption difficult for the company to overlook.
Microsoft's first customer is itself
https://x.com/CompleteSkeptic/status/2108594987177021737?ref_src=twsrc%5Etfw
Microsoft Chairman and CEO Satya Nadella announced Decision-1 on Friday via X, stating, "We're already testing it across Microsoft," a claim supported by four internal teams. Xbox Research deployed it to categorize over 10,000 player feedback entries, the Copilot team utilized it for evaluating chat and agent outputs, on-call engineers leveraged it to gather context during active incidents, and Microsoft Discovery applied it to assess experiments before agent replanning.
Microsoft's internal testing revealed Decision-1 operates more than 14 times faster than GPT-6 Sol at significantly lower cost for Xbox operations, and demonstrates 46 times greater consistency within Discovery workflows. The model may serve a broader strategic purpose: Microsoft disclosed Wednesday that GitHub Copilot will soon determine whether tasks execute locally or route to cloud-scale models, though specifics about what Copilot transmits to the cloud remain undisclosed.
https://x.com/satyanadella/status/2108627923888754862?ref_src=twsrc%5Etfw
Model routing appears among Decision-1's documented applications, yet Microsoft has not clarified whether the model will govern Copilot's routing decisions. Given routing requirements spanning Copilot, GitHub, and Xbox, Microsoft possesses strong motivation to manage these operations internally.
Decision models became cheap
Cognition's vice president of engineering, Jared Palmer, spent approximately $95 in Modal H100 compute resources to adapt his open-source Kev models to Qwen3.5, while Cloudflare constructed Clef-flash using the identical Qwen3.5-9B foundation that Microsoft selected.
Jev's pricing has become the industry standard: Palmer lists Kev-4B on OpenRouter at $0.042 per million input tokens, Perplexity charges $0.02, and OpenAI's rate exceeds Jev's by more than double. At these rates, individual decision call revenue remains marginal, shifting Microsoft's primary opportunity toward retaining agent traffic—including the generative operations surrounding each decision—within Foundry. The OpenRouter listing may additionally attract developers outside the Azure ecosystem into Microsoft's platform.
Achint Srivastava, vice president of software engineering in Microsoft's Office of the CTO, presented Decision-1 as enabling decision-making capabilities for existing applications, agents and workflows "in a secure, trusted environment."
Text only, no weights
This reveals Decision-1's limitations relative to competitors. Its Foundry documentation specifies support for up to 32,768 text tokens and JSON output, but excludes image processing—a capability both OpenAI's Decisions API and Cloudflare's Clef, which incorporates a vision encoder, provide.
Cloudflare released Clef under Apache 2.0 licensing, whereas Microsoft has not yet announced open weights. AWS, Upstage, and Ollama have embraced TypeSafe's System One API as a standardized interface across the category, but Microsoft has not confirmed whether Decision-1 achieves full compatibility, despite sample code in Foundry referencing a /systemone endpoint.
Calibration under adversarial pressure
Microsoft states that Decision-1's probabilities are calibrated, meaning a 90% prediction should prove correct approximately nine times in 10 across representative datasets. However, research on Jev demonstrates how substantially confident scores can deteriorate when inputs are crafted to mislead. Microsoft's own Foundry documentation recommends customers validate calibration performance against their proprietary data.
The JevOut preprint, authored by USC computer science researcher Zixiang Xu and collaborators, revealed that brief, plausible additions to context reversed 312 of Jev's 508 initially correct decisions. In 229 instances, Jev assigned at least 70% probability to the incorrect answer. Three alternative scoring systems exhibited flip rates between 64.9% and 73.2% under identical testing conditions.
Microsoft subjected Decision-1 to eight perturbation categories, including reordered options and rephrased descriptions, which altered approximately 1.3% of responses on average. JevOut did not evaluate Decision-1, leaving uncertainty about the model's resilience against comparable attacks. Microsoft has additionally not confirmed full System One API support among competitors, raising unresolved questions regarding interoperability and the appropriate confidence level to assign to its probability scores.