Microsoft-Decision-1 is available in Foundry to score fixed answer options, making classification a low-cost model call.Availability That capability has a testable limit: if a support system’s options do not cover the user’s problem, faster classification can still send that person to the wrong place. Savings on inference need to outweigh the cost of transfers and follow-up questions.
The official announcement is dated 2026-10-09, without a publication time or timezone. RuntimeWire published its report at 15:21 US Central time that day, equivalent to 04:21 on October 10 in Taipei. This article’s cutoff is 15:50 Taipei time. It uses a previous-day release within the 48-hour window; the reporting date crossing midnight does not establish that access began today.Announcement Report timing
Microsoft post-trained Qwen3.5-9B for this model. An application supplies context, a question and predefined answers, then receives probabilities for the options. The model does not generate explanations and is unsuitable for open-ended questions.Model card Developers can therefore assign routing to a dedicated model while leaving response writing to a generative model.
The published price is $0.042 per million input tokens, with free output.Pricing Foundry’s launch post describes public preview, whereas the catalog labels the model GA. Those lifecycle descriptions conflict. Access is documented, but the documents do not establish when a transition to general availability occurred.Launch status Catalog status
Microsoft reports the highest accuracy in its comparison of 36 benchmarks covering nearly 150,000 questions, with benchmarks kept blind from training. These are vendor results, not an outcome established for every customer’s workload.Evaluation Xbox Research used the model to assign more than 10,000 feedback items to researcher-defined themes. That internal case demonstrates a use, without establishing ordinary customers’ long-term results.Internal testing
Decide which needs can share a category
The Xbox example makes the product choice concrete: researchers already know which themes they want to compare, and a model can repeatedly process large volumes of feedback. If the purpose is to discover previously unanticipated problems, restricting answers to existing themes can instead place a new need in the nearest old category. This follows from the fixed-choice mechanism; it is not evidence that Xbox actually missed such cases.
Suppose a user requests a return while reporting a product safety concern. If the system offers only “refunds” or “technical support,” a single choice may lose the other need. I would assess the two questions separately and preserve the original message for subsequent handling, rather than require one label to summarize everything. Both needs can be valid at once; better classification accuracy cannot repair an inappropriate single-choice design.
The model card permits a “cannot tell” option and assigns thresholds and escalation to the integrating application. It also excludes using the model alone for consequential decisions about people.Usage limits Low-cost scoring is therefore a plausible fit for clearly defined, reviewable routing. Work that repeatedly requires discovering new options or obtaining outside knowledge exceeds the scope this release addresses.
Lower pricing could also change how software allocates work between models. A product that currently combines classification and reply generation in one call could separate them, allowing category rules to change without also changing how replies are written. In return, the team must maintain the mapping from categories to subsequent handling. That cost depends on how often the product changes; a token price cannot settle it for every team.
I see the stronger fit in repetitive work with stable options, rather than treating low pricing as a reason for wholesale replacement. Useful measures are the number of transfers before a user reaches the correct handler and the proportion of requests without a suitable option. If call costs fall but those measures worsen, the questions and options need revision. Finishing classification faster would not yet establish a better product.
The cover reuses this site’s Microsoft Build developer-event artwork, not a Decision-1 interface screenshot. Downloading the official event image failed because DNS resolution was unavailable.
Sources: