A quiet category of AI model has gone from niche to crowded in about three weeks. “Decision models” do not write essays or code. They take a question and a fixed list of possible answers, then return a score for each option. Since late September 2026, OpenAI, Microsoft and Liquid AI have each shipped one, and the pitch is the same: the routine sorting inside an AI agent should not cost as much as the thinking.
What a decision model does
InfoWorld’s description of Microsoft’s entry is the clearest. A general-purpose language model generates free-form text, while a decision model chooses from a predefined set of answers and returns a probability for each in one pass. Software can then act on those numbers directly: send this support ticket to billing, mark this output as meeting the rubric, escalate this alert or not.
That matters because agents spend a lot of their time on such choices. Forrester analyst Charlie Dai told InfoWorld that routing, classification and approval checks “often represent a large share of agent interactions.” If a small, fast model can handle those steps, the expensive model is only called when real reasoning is needed.
Three launches, three price points
OpenAI announced a Decisions API at its DevDay event in San Francisco, in an article from The Decoder dated September 29, 2026. It is built on GPT-6 Luna, OpenAI’s low-cost model. Developers define a question with limited answers and pass in text or images; use cases include content classification, request routing and deciding an agent’s next step. OpenAI says it is ten times faster than asking Luna the same thing through the regular API, with a chart showing 150 milliseconds against 1.6 seconds. It launched as a limited preview, with broader availability promised in the following days. A release tracker lists a public beta by October 8, though that detail comes from a single secondary source.
Microsoft’s model, called Microsoft-Decision-1, is a 9-billion-parameter system built by post-training Alibaba’s Qwen3.5-9B. It is available on Microsoft Foundry and OpenRouter and costs 4.2 cents per million input tokens, with no charge for output tokens. Microsoft’s own example is an incident-prioritization request of 86 input tokens and 2 output tokens that costs roughly $0.0000036, meaning a dollar would cover more than 275,000 such decisions. Microsoft says it is about 2.5 times faster than the runner-up and 35 times faster than GPT-6 Sol across 36 benchmarks covering nearly 150,000 held-out questions. Those claims have not been independently verified.
Liquid AI took a different route with its d1 family, released as open-weight models on Hugging Face: d1-3B, which accepts text and images, and an experimental d1-omni-600M that handles text plus images or audio. Liquid’s post reports single-question latency of 8 milliseconds on an RTX 4090, 30 milliseconds on an Apple M5 Pro, and 50 milliseconds on a Jetson Orin Nano, and says d1-3B scores 48.57 on a public decision benchmark, ahead of every model under 10 billion parameters. The company demonstrates live camera uses such as content moderation and gesture control. The post does not name a specific license, so anyone planning commercial use should read the model cards first.
The catch is explanation
The same design that makes these models cheap also makes them opaque. A decision model returns scores, not reasoning. Kanerika’s Amit Kumar Jena, quoted by InfoWorld, warned that “a precise-looking probability can give an uncertain decision more authority than it deserves.” He also noted that, per Microsoft’s own documentation, results can shift with the wording of the question.
Dai’s recommendations are sensible and cheap to follow: test calibration, control the wording and thresholds, keep audit logs, and keep a human in the loop for high-impact decisions. For anything touching credit, hiring, insurance or health, a number without a rationale is a poor basis for a decision that someone may later need to defend.
There is also a measurement problem. Vendors quote per-token prices, but what a team actually pays is cost per completed workflow, including retries and fallbacks to a bigger model when confidence is low. Jena recommended measuring exactly that rather than trusting benchmark-derived savings. A model that is 35 times faster on paper can still lose its advantage if a third of its answers get escalated.
What This Means for Philippine Founders
For startups running customer support, KYC checks, content moderation or lead scoring, the routing layer is the obvious place to try this. A small team can keep its current large model for hard cases and put a decision model in front for sorting, then compare monthly spend and accuracy on a few hundred real examples, including messy Taglish messages. Most published benchmarks are English-centric, and whether any of these models handle Filipino or code-switched text well has not been established in the sources reviewed here, so local testing matters more than the vendor charts.
The open-weight and small-model angle may matter more than the price. Liquid’s d1 models are designed to run on modest hardware, which suggests workable on-device or on-premise classification for teams with data-residency concerns or unreliable connectivity. Whether that holds for a given product depends on the license terms and on accuracy in the languages customers actually use. Founders should treat the numbers above as vendor claims, run a small pilot, and keep a human review path for any decision with real consequences for a customer.
Share this article