Decision models after Jev: which one fits which job

On October 20 a hands-on test of Perplexity’s pplx-decider-v1.2 gave the model one customer message about a failing API integration and got three answers back in a single pass: the support team at 99.8% confidence, blocked yes at 98.4%, and a frustration score of 1.28 on a four point scale. No paragraph, no tokens generated. That is the shape of the whole category. I would use a decision model in all AI workflows, in classification and in automation, because a huge part of AI usage is workflow automation, and the speed and the token efficiency open realtime scenarios that were not reachable before.

The word decision no longer covers all of these models. Cloudflare wrote of its new Clairic that “Until now, every System One model has been a decision model.” Clairic is the first extraction model and answers differently: a number with an 80% range, a name or an order ID as a span with character offsets, not probabilities over options you wrote in advance. Cloudflare’s own examples show why. What is the invoice total, when was this delivered, how many days late is it. None of those can be listed as options, so no decision model can score them.

One pass carries the cost

pplx-decider-v1.2 is a 27 billion parameter model on the Qwen3.8-27B backbone under Apache 2.0, and it needs roughly 28 to 29 GB of VRAM. The release lists three changes from v1.1, and the first is joint reasoning across multiple questions in a single pass. The single-pass change is the right bet. Three answers on one ticket in one call is what routing actually costs, and v1.1 paying for three calls was the real tax.

The same arithmetic repeats across the pages. Liquid bills d1-ultra at $0.042 per million input tokens with no output tokens at all, and a 1024 by 1024 image counts as 1,536 tokens. Microsoft measured a P50 latency about 35 times faster than GPT-6 Sol, and does the sum that matters: 100 milliseconds added to each of 20 sequential decisions adds two seconds to the workflow. Clef-omni returns a text decision in about 130 ms at the median and scores a 21 second video with sound in about 1.5 seconds, in one call.

Modality moves faster than accuracy

Clef-omni takes audio in wav or mp3 and video in mp4 or webm alongside text and images, so there is no chain of transcribe then classify. It is built on a Qwen3-Omni-30B-A3B mixture of experts foundation with the speech output parts discarded. Clairic accepts text plus up to four images. Jev’s image support sits behind a header flag, TypeSafe-Danger-Zone-Images: true, and TypeSafe says of its own models that “all models (Jev included) perform worse on multimodality over pure text.” The Perplexity test backs the caution: 65.9% toward yes on a seatbelt worn incorrectly, and 89.2% on a seatbelt judged against Indian road rules.

A probability needs a threshold

Liquid says it trained d1-ultra specifically against overconfidence, because most decision models are overconfident. A probability is only worth the threshold behind it. Liquid trained the model against overconfidence, but someone still has to pick the number that triggers a return on a damage claim. Microsoft publishes a number in the same spirit: the model changes its decision on 1.3% of perturbations on average, with zero flips when option descriptions are paraphrased or the options are shuffled. For a router that number says more than a leaderboard place.

I am not sure about accuracy. On paper it should be really good, but I don’t trust benchmarks, I want to see it on the field. The pages give reasons to stay that way. Perplexity states that some reported gains came from training on official benchmark splits, TypeSafe writes that it dislikes public benchmarks, and the Microsoft comparison of 36 benchmarks and nearly 150,000 questions is Microsoft’s own. There is a second dependency. Microsoft built Decision-1 by post-training Qwen3.5-9B and will soon rebase it on MAI and OpenAI. A router whose scoring shifts under a base model swap is a dependency I would pin, not adopt.

The unglamorous work comes first

Clairic’s list, invoice total, delivery date, days late, repair cost, traffic over the next 24 hours, is the work that happens before any decision gets made. Extraction from unstructured input is the unglamorous job that makes agentic workflows actually useful in commerce, and it is the class here that returns a typed value or null instead of a probability. The other lesson is about flags. Putting the dreamer features behind a danger zone header is the honest version of what most vendors ship quietly, and I have inherited a platform where that flag was already on in production. The scoring is cheap now. The threshold and the pinned version are still a person’s call.