Microsoft, Baseten and Cloudflare All Shipped Decision Models in October

I wrote about Jev in September because it looked like a new category rather than another chatbot. TypeSafe AI's pitch was a model that answers typed questions with calibrated probabilities and generates no tokens at all, so it structurally can't return an option you didn't give it. I didn't expect three more serious entrants in the following six weeks, but that's what happened.

On October 9, Microsoft shipped Microsoft-Decision-1, a post-trained Qwen3.5-9B that scores yes/no judgments, multiple-choice options and ratings against calibrated probabilities instead of generating text. On October 19, Cloudflare previewed Clairic, which it calls the first of a new "System One" class of extraction models: it pulls typed values (numbers, dates, estimates, forecasts) out of unstructured input rather than answering predefined questions. Two days after that, Baseten's research arm Base Labs released Mammoth-1 and Elephant-1, a 27B and a 9B decision model fine-tuned from Qwen3.5. Both currently top the Jev Decision Index, a Hugging Face leaderboard that happens to be named after the model that started this.

The common shape

Every one of these charges for input tokens only, because there's little or no output to bill for. Every one returns a probability or confidence value instead of a bare answer, so you can set a threshold and escalate the uncertain cases instead of trusting a single number. And every one is pitched at the same job: the high-volume, low-ambiguity decision that a lot of production systems currently burn a full generative call on. Which queue does this ticket belong to? Does this response violate policy? Should this tool call be allowed to run? None of that needs a model that can write a paragraph.

Where they actually differ

Microsoft-Decision-1 and Mammoth-1/Elephant-1 are the same basic move. Take an open-weight generative model (Qwen3.5 in both cases) and post-train it to emit probabilities over a closed answer set instead of free text. That's a constraint applied on top of a generative backbone. Jev's pitch is different. It's a parallel sampler that reads the input once and scores every question in the same forward pass, with no autoregressive generation step to constrain in the first place. Whether that distinction holds up under real traffic is a separate question from whether it's marketed that way. I haven't seen anyone test it directly.

Clairic is the odd one out here and probably shouldn't be grouped with the other three by job, only by its generation-free architecture. It's not answering questions you wrote against a fixed option set. It's extracting or computing a typed value, a date, a duration, an estimate, out of messy input. That's closer to structured data extraction than to a routing or gatekeeping decision. Cloudflare's own demo, a damaged-claims pipeline, finished in 19.3 seconds against 61.7 seconds for an all-LLM baseline, at 49% lower cost. That's a believable number for extraction. It isn't really comparable to a yes/no gatekeeper call though.

The numbers are still vendor numbers

Microsoft reports 83.5% average accuracy across 36 benchmarks and 85ms p50 latency. That's Microsoft's own benchmark suite, not an independent one. Cloudflare's 49% figure comes from one internal pipeline demo. The Jev Decision Index is the closest thing to a neutral signal here, since it's run on Hugging Face with published rules against prompt tuning, truncation and option filtering. Mammoth-1 currently leads it. That's still one leaderboard measuring one kind of accuracy. It doesn't tell you about calibration quality or about how any of these models fail on a question their own backbone is already weak at.

What I'd actually do with this

A market with four comparable products beats a market with one. A single vendor's roadmap or pricing change can no longer silently decide your architecture for you. The practical move hasn't changed from the one I'd have made with Jev alone: write the integration behind your own interface, with your own confidence threshold and escalation logic, so swapping the backend is a config change rather than a rewrite. What's new is that the choice of backend is now worth actually making. None of the four vendors has published enough independent evidence yet to make that choice for you.