· via TechCrunch
Microsoft's Decision-1.1 arrives as decision models take a quarter of Vercel AI Gateway traffic
More than one in four requests through Vercel's AI model router now go to decision models, a class that returns structured answers instead of text. Microsoft's Decision-1.1 has just landed on the service.

Decision models capture a quarter of Vercel traffic
More than one in four requests running through Vercel's AI model router in October went to decision models rather than traditional language models, according to TechCrunch. That is a striking share of production work for a model category that, as TechCrunch points out, only emerged about a month earlier.
Decision models are language models that generate no output tokens at all. Instead, they apply their general capability to choose quickly and cheaply among preset options, while reporting how confident they are in the choice. According to TechCrunch, this suits production environments where a fast, calibrated decision — routing a ticket, picking the next step in a workflow — matters more than the extended reasoning of a slower and more expensive frontier model.
Microsoft's Decision-1.1 lands on AI Gateway
On the same day, Vercel announced on its blog that Decision-1.1 from Microsoft AI is now available on AI Gateway. The model returns structured answers instead of free-form text, aimed at tasks such as routing support requests, selecting the next step in a workflow, or grading AI responses.
Vercel documents three question types: Boolean, which returns the probability that a condition is true; Choice, which returns a selected category along with probabilities for every option; and Score, which returns a numerical rating on a scale the developer defines. Multiple questions can share the same input in a single request.
Migration is deliberately simple. Decision-1.1 accepts the same request format as Decision-1, and developers only need to change the model ID to microsoft/decision-1.1. It can be called through the AI SDK, the OpenAI-compatible Decisions API, or the AI CLI, with Vercel specifying ai 7.0.128 or later for TypeScript and OpenAI Python SDK 3.26.0 or later for Python. Decision requests appear in request logs, count toward budgets, and support decision fallbacks.
The gateway debut follows TechCrunch's report that Microsoft CEO Satya Nadella said on Tuesday that the model was being rolled out for wide use on Azure immediately.
A crowded field built in weeks
According to TechCrunch, the category's breakout player is TypeSafe, which reached a $7.5 billion valuation after releasing its Jev family of models. Jev still dominates usage where metrics are visible: OpenRouter alone handled more than 1.5 billion queries to TypeSafe's models last week.
The barrier to entry, however, is low. TechCrunch reports that dozens of decision models appeared in Jev's wake — one Amazon engineer built one over a weekend — and every hyperscaler now fields its own option.
Startups are piling in as well. Sqwish, a U.K. company behind a decision model called kettei fat, announced a $3 million pre-seed round. Co-founders Ushnish Sengupta and Federica Freddi, who met as Cambridge graduate students, had been building small LLMs and realized those architectures could be repurposed once Jev arrived.
Data and tuning over benchmarks
TechCrunch reports that data is emerging as the deciding factor in the category. TypeSafe CEO Diogo Almeida said the hardest part of building Jev was finely calibrating its judgement with statistically adjusted synthetic data. Sqwish also uses synthetic data, but differentiates with built-in post-training tools that improve the model's judgement using a customer's own production data.
Sengupta told TechCrunch that a benchmark is "more of an invitation to try the model out," with continual tuning the reason customers stay. Microsoft engineering VP Achint Srivastava, who led the Decision-1.1 project, argued that "benchmaxed" models fall flat under real enterprise usage, and claimed that Decision-1.1 and Jev are among the few that perform strongly across accuracy, reliability, speed and cost, even on held-out benchmarks.
Freddi predicts that "System One" intelligence like decision models will increasingly work in concert with "System Two" frontier LLMs — and she frames even that narrower slice as a $10 billion to $50 billion opportunity, big enough for a startup while remaining a side project for giants.
Why it matters
A model class that barely existed a month ago already carries over a quarter of the traffic on a major routing service. That is a signal about how AI workloads are splitting in production: instead of one generative model doing everything, cheap, fast, structured judgement calls are being carved off and handled by models purpose-built for decisions.
For developers, structured answers remove a persistent pain point — parsing free-form text out of an LLM — and Vercel's drop-in model ID swap makes trying the new class nearly free. For the industry, the speed at which the field commoditized suggests the durable advantage is not the model itself but the data and continual tuning behind it, which is where both TypeSafe and Sqwish are placing their bets.
- #decision-models
- #vercel
- #microsoft
- #ai-gateway
- #llms
- #machine-learning