Microsoft Decision-1 launched on 9 October 2026, and it does one job. It scores a fixed set of options and hands back a probability, not a paragraph. It's cheap, it's fast, and it's aimed at the routing and verification steps buried inside every chatbot and automation we build. If you run AI workflows for your business, this is worth two minutes of your coffee.
What Exactly Is Microsoft Decision-1?#
Decision-1 doesn't generate text. Give it a situation and a defined set of choices, and it returns a calibrated probability for each one as structured JSON, covering yes/no calls, multiple choice, star ratings, and rubric grading. That's the whole model.
Satya Nadella called it "our new model for fast decision-making" when Microsoft announced Decision-1 on 9 October 2026, built by post-training Alibaba's open-weight Qwen3.5-9B. It takes text up to 32,768 tokens and returns no explanations, just the probability. It's live in Microsoft Foundry now, priced at $0.042 per million input tokens with output tokens free, since there's nothing generated to charge for.
On Microsoft's own 36-benchmark test, Decision-1 scored highest for accuracy, and its typical response landed 35 times faster than GPT-6 Sol. Runtimewire reported that Xbox Research ran it against more than 10,000 feedback items at roughly 200 times lower cost than an LLM-based judge.
Does This Change Your Automation Setup?#
If every AI step in your chatbot or lead pipeline runs through a full language model, even to answer a plain yes-or-no question, you're paying full-generation pricing for a decision that never needed generated text at all. That's the gap Decision-1 is built to close.
Decision-1 isn't the first attempt at this either. OpenAI shipped something similar with its Decisions API in September, and the economics argument will sound familiar if you've been watching WhatsApp bot pricing shift this year too.
At WebEpex, we build WhatsApp and web chatbots for clients like Lionaire Concierge on Voiceflow, and the step between "the knowledge base can answer this" and "this needs a human" is exactly the kind of call that's always eaten into the AI budget on builds like that.
If you're not running any AI automation yet, or your bot is one scripted flow with no judgment calls in it, this doesn't touch you. Skip to the close.
If you've got a chatbot, a lead router, or an agent deciding "escalate or don't" more than a few hundred times a day, this is the layer worth rebuilding around. The savings are small at low volume. They add up once you're past a few thousand decisions a month, which is where most of the automations we build already sit.
Is It Worth Switching To Right Now?#
My honest read is that the idea is sound, but every number behind it is Microsoft's own. The speed and cost claims are probably directionally true. I wouldn't bet a client migration on the exact multipliers today.
The upside is real. A model built only to score fixed options should still beat a full chat model reasoning through the same choice in prose.
Here's the honest catch. Neither TestingCatalog nor Runtimewire could find published datasets or latency test conditions to check the comparisons independently. Treat "35 times faster" as a Microsoft claim, not an audited fact, until someone outside Redmond runs the same test.
There's a real limit worth naming too. Decision-1 doesn't explain itself. It hands back a probability, not a reason, so you lose the "why did it pick this" trail a full model gives you for free. And Microsoft's own documentation says it shouldn't be the sole decision-maker on anything touching employment, credit, housing, healthcare, or legal rights. That's the right caution, and worth repeating to anyone building HR or lending bots on top of it.
How We're Already Testing It Against Our Routing Layer#
At WebEpex, we pulled Decision-1's pricing and benchmark claims apart within a day of Microsoft's announcement and mapped them against what our own routing step costs per thousand decisions in the n8n workflows we build for clients.
We've run small classification models in that exact spot for a while now. Not because it's clever. Because paying full LLM rates to answer "is this a complaint, yes or no" stopped making sense once a client's message volume climbed past a few hundred a day.
We're not moving a live client flow onto a model that's two days old. That's not how we handle real customer traffic. I'd rather be slow and right than fast and sorry. What we have done is add Decision-1 to the shortlist we test against on the next routing-heavy build, alongside the classification setups we already trust.
We build chatbots on Voiceflow and automation flows in n8n for clients across the GCC, European, US, Canadian, and Indian markets we serve. A model this cheap, one you can afford to run on every message instead of just the expensive ones, is worth the half-day it takes to benchmark properly.
What I'd Tell A Client This Week#
If you asked me this week, I'd say it straight. If your automation makes fewer than a couple hundred AI-assisted decisions a day, don't touch anything. The savings won't justify the engineering time.
If you're past that, pull up your current flow and find every spot where you're calling a full model just to get back a yes, a no, or a simple priority score, because that's your shortlist, not the whole pipeline. Don't rip out anything live yet. Test the swap on one low-stakes decision point, watch it for a week, and only let it near a real customer after that.
We're making that same slow-and-right call on our own chatbot and automation builds.
If you want a second pair of eyes on where your own setup is burning model calls on decisions that don't need generated text, send me what you're running and I'll tell you straight, two minutes, no pitch. [cal.com/webepex/growth-review]
