Microsoft Launches Microsoft-Decision-1, a Model 35x Faster Than GPT-6 Sol

Microsoft has introduced Microsoft-Decision-1, a specialized model that does not hold conversations and does not write text. Its job is to evaluate predefined answer or action options and pick the right one. According to the company, its response latency in structured decision-making tests is 35 times lower than that of OpenAI's GPT-6 Sol. It is intended for classifying information, verifying outputs, and managing workflows.

What the model actually does

A typical language model receives a prompt and generates a response token by token. Microsoft-Decision-1 is built for a different task: it takes a context and a list of permissible outcomes as input, and the output is a choice. For example: which category a request belongs to, whether to pass a result further down the chain or send it back for rework, which step to run next.

Operations like these are everywhere in enterprise systems. Teams have often handed them to a large general-purpose model, even though the answer fits in a single word or an option number. Decision-1 targets exactly that scenario.

What the numbers show

The published data names only a few parameters:

• model class: specialized, evaluates given options, does not conduct dialogue;
• main metric: response latency 35 times lower than OpenAI's GPT-6 Sol;
• comparison conditions: structured decision-making tests;
• tasks: classification of information, verification of outputs, workflow management;
• source of the figures: Microsoft's own data.

The announcement does not include the parameter count, pricing, access format, or accuracy figures. These matter most for any deployment decision. Speed guarantees nothing without quality of choice.

Where the 35x gap comes from and how to read it

A plausible explanation lies in the nature of the task itself. If a model does not need to produce extended text and only has to score a finite number of options, the computation is shorter. A large generative model has to spend compute on versatility that this scenario does not require. This is an inference from the logic of the task, not an architecture disclosed by Microsoft.

That is why the comparison deserves caveats. A narrow model is tested on the task it was built for, while a general-purpose flagship model serves as the baseline. A latency gap is expected in that setup. Open questions remain: how closely the decisions of the two models match, how Decision-1 behaves on data outside the test set, and what happens when the list contains no suitable option. There are no independent measurements yet.

The direction itself is telling. The market is moving away from one big model for everything toward a combination: a large model plans and reasons, while compact specialists quickly handle routine decision points. For agentic systems this is critical, because an agent runs dozens of small checks per launch, and the latency of each one adds up.

Where this is useful in practice

The first candidates for use are obvious:

• routing requests and documents by category;
• checking another model's outputs before they reach the user: whether a response complies with the format and rules;
• choosing the next step in an automated process where the set of actions is known in advance.

Teams building agent chains should look through their logs and find calls to a large model where the answer boils down to a choice among a few options. Those are the calls worth testing on a specialized model. Three parameters should be compared: latency, the share of matches with reference labeling, and cost per call. Until Microsoft publishes pricing and access terms, it is too early to count the savings. But the pitch is clear: fast decisions in large systems should not have to cost as much as a full dialogue with a flagship.