What happened
On October 6, 2026, OpenAI put its Decisions API in public beta. Its documentation says it evaluates text, images, or both and returns typed answers: the probability that a condition is true, a choice from a fixed set, or a score against a rubric. OpenAI says it answers about 10 times faster than its Responses API, that gpt-6-luna is the only model it runs on for now, and that it expects general availability in the coming weeks. On October 9, Microsoft launched Microsoft-Decision-1 in Microsoft Foundry, with OpenRouter access to follow. Microsoft says it post-trained the open-weight Qwen3.5-9B model for the job and will soon rebase it on Microsoft AI and OpenAI models. Microsoft lists input at $0.042 per million tokens, with output tokens free. Its own documentation describes uses such as triaging support tickets, prioritizing incidents, filtering content, and picking which model or tool should handle a request. Also on October 9, Cloudflare cut the price of its Clef-flash decision model from $0.09 to $0.038 per million input tokens. In the same post it said the hosted version's context window is now 24k, instead of the 64k it had advertised. Cloudflare says only 0.24% of requests exceed 24k, and that the downloadable weights are unchanged. The Register reported on October 10 that more than 100 such models are now competing, and that OpenAI, Cloudflare, Perplexity, Snowflake, and others have all entered the category in the weeks since TypeSafe AI announced a model called Jev.
Why it matters for business owners
Most AI talk is about models that write, summarize, or chat. A lot of the daily work in a business is something else: decide which team gets this ticket, flag which invoice looks wrong, rate how urgent this email is, check whether a photo shows damage, or choose which of two tools should handle a request. Those are sorting jobs. If you run them through a general chat model, you may be paying for writing ability you never use, and waiting for answers you do not need written out. A decision model is built for exactly that shape of job, and the vendors are pricing it very low. Microsoft also frames the bigger use as a check inside agent workflows: score the agent's proposed next step, then continue, stop, retry, or hand off to a person. For an owner thinking about agents, that is the more interesting idea. The cheap model can act as a gate in front of the expensive one.
What owners should not misunderstand
Do not treat the vendor numbers as your numbers. Microsoft says its model had the highest accuracy in its own 36-benchmark comparison, and that internal teams found it competitive with larger models on tasks such as labeling survey feedback. Those are Microsoft's tests, run on Microsoft's data. OpenAI's speed claim is OpenAI's. Your tickets, invoices, and customers are different. Do not read "decision model" as "cannot be wrong." The Register notes that decision models are sometimes described as not hallucinating the way chat models do, and then adds that they can make errors, and that researchers are already studying how those errors might grow. Microsoft itself reports that its model changed its answer on 1.3% of the test variations it tried. That is a small number on a test. On a thousand real cases a day, it is still real decisions that go the other way. Do not assume a confidence score is a guarantee. Microsoft says a 90% prediction should be right about nine times in ten on representative cases. The words that matter are "representative cases." Whether your cases are representative is something only you can check. Do not treat a low price as a fixed term. Cloudflare's cut came with a smaller context window on the hosted version, announced in the same post. OpenAI's API is still in public beta, with general availability expected in the coming weeks, so terms can change. Microsoft says the base model underneath will change. A cheap price and a short list of conditions can move together, so read what changed, not just the price. And do not assume any of this replaces a chat model. Microsoft's own documentation says to use a generative model when you need to create, summarize, rewrite, or explain, and suggests using the two together.
The operational lesson
Before you pick a model, split the workflow into steps and label each one as either sorting or writing. Sorting steps have a fixed set of answers: a team, a priority, yes or no, a score. Writing steps produce text a person will read. Sorting steps are the candidates for a smaller, cheaper, faster tool. Writing steps usually are not. The mistake is paying chat-model prices and waiting chat-model time for a task that is really a label. The second part of the lesson is how you decide whether a sorting tool is good enough. The question is not how it scores on a leaderboard. It is how often it is wrong on your cases, how costly each kind of mistake is, and who sees the cases it is unsure about. A wrong ticket route costs a few minutes. A wrong decision to approve a refund, flag a customer, or clear a document costs more. The same tool can be right for one and wrong for the other. The third part is ownership of the threshold. These tools return a probability, and your software or your team decides what to do with it. Someone in your company has to own the line between "act automatically" and "send to a person."
What a serious business should do next
Pick one workflow where people sort things all day, such as support tickets, incoming leads, or invoice exceptions. List the possible answers. If you cannot list them, it is not a sorting job yet. Pull a few hundred real past cases that a person already labeled. Hold them back as your test set. When you try any vendor's model, including a free trial, score it against those labels and write down the error rate and the kind of errors. Do not use the vendor's sample data as the test. Decide in advance what happens to low-confidence cases. The simplest rule is to send them to a person. Start with the model suggesting and a person approving, and loosen it only after the error rate holds up over a few weeks. Keep the spend and the terms in view. Note the price, the context limit, and whether the product is in beta. Check any data handling terms before you send customer information to a new endpoint, and confirm where the data is processed. If the process is rare, small, or already fast, this may not be worth doing yet. A decision model helps most where the same kind of judgment repeats hundreds of times a week. Nothing here is legal or financial advice.
The Atlacis view
Atlacis helps owners slow down before an AI decision, map the workflow, and choose the simplest tool that fits the step. Decision models are a good reminder that the right tool is often smaller than the one in the headlines, and that the proof is a test on your own cases, not a launch post. If you are weighing where AI belongs in a workflow and want help separating the sorting steps from the writing steps, and setting a sensible test before you commit, that is a good conversation to have early.
The short version
- Between October 6 and 9, 2026, OpenAI, Microsoft, and Cloudflare launched decision models: small models that return a probability, a choice from a fixed list, or a score instead of written text.
- Many business AI tasks are sorting, not writing. Label each step in a workflow as sorting or writing before choosing a model.
- Every accuracy and speed number in these launches comes from the vendor. Test on a few hundred of your own labeled cases before relying on one.
- A low price can come with changed terms. Cloudflare cut its hosted context window from 64k to 24k alongside its price cut, and OpenAI's API is still in beta.
- Assign someone to own the line between acting automatically and sending a case to a person. Start with a person approving.
Where ATLACIS can help
Sources
- OpenAI: Decisions guide, OpenAI API documentation
- Microsoft Command Line: Introducing Microsoft-Decision-1, our model for fast decision-making (October 9, 2026)
- Microsoft Learn: Deploy and use Microsoft-Decision-1 in Microsoft Foundry
- Cloudflare Blog: Introducing Clef-omni with full multimodality, plus a faster Clef and a cheaper Clef-flash (October 9, 2026)
- The Register: Microsoft leans on open weight model from Chinese AI lab to challenge Jev (October 10, 2026)