Decision models
Paper
A decision model does not write text. You give it a piece of text and a few questions, and it answers each one with a probability: which queue, how urgent, yes or no.
Jev from TypeSafe started the class in September 2026, and in October d1-ultra from Liquid AI took the top of the Artificial Analysis index. This is a note on what these models are, where they are already used, and how far they have really caught up with ordinary LLMs.
What it is
An ordinary LLM answers by writing: it produces one token after another until the answer is finished. Asking it to classify a text means asking it to write the word "billing" and hoping the word is one of the allowed ones.
A decision model skips the writing. It reads the text together with typed questions and returns a probability for every option in a single pass. There is no prose and no reasoning to read, only the distribution. Jev, d1 and the other models of this kind are built around the same few question types:
- Choice
- Pick one option from a list you wrote. Returns a probability for each option.
- Score
- Place the input on an ordered scale you describe, for example urgency from 1 to 5. Returns a position on that scale.
- Noul
- A yes or no question. Returns one number from 0 to 1. Noul is Jev's name for it, in d1 it is a plain yes/no question.
- Box
- Only in Jev 1.14 and only for images: returns a bounding box around the thing you asked for.
TypeSafe says the probabilities are calibrated: when the model says 0.8, it should be right about 80% of the time across many answers. That is the vendor's claim, and one independent write-up warns that confidence shows how concentrated an answer is, not whether it is correct.
Who makes them
The class is a few weeks old, so the list is short and changes fast. By "Jev-style" I mean any model of this kind: typed questions in, calibrated probabilities out, no generated text. These are the ones I could place:
- 01
Jev
TypeSafe. Released on 15 September 2026, and Jev 1.14 followed on 14 October with image input. Proprietary, no published weights.
- 02
d1 and d1-ultra
Liquid AI. d1-ultra (15 October 2026) is the strongest of the family and takes text and images. The family also has open-weight d1-3B (3.1B parameters) and d1-omni-600M (587M, can take short audio).
- 03
The rest of the index
Microsoft-Decision-1, Decider v1.1 27B from Perplexity, Mercury Decide from Inception, Celeris-1 Decision, GLIDE from Fastino, Clef and Clef-Omni from Cloudflare, and GPT-6 Luna (Decisions) from OpenAI.
Most of them are hosted and closed. Only the small d1 models publish their weights.
What they are for
For classifying text. That is the whole job, and a lot of software is made of it: where this message goes, whether this request is allowed, whether this answer follows from its source.
An LLM can do it, but it pays for things the job does not need: it generates tokens, which costs time and money, and it is a model big enough to write essays. A decision model only reads. Its answer is a few dozen tokens, and the hosted ones do not bill output at all.
They are also small. d1-omni-600M has 587M parameters, d1-3B has 3.1B, and Decider v1.1 27B is the largest I found with a public size.
- $0.042per million input tokens, output free. The listed price of Jev and of d1-ultra.
- $0.10 and $0.50per million input and output tokens of Claude Haiku 5.5, for prompts under 100K tokens.
Route a ticket: queue, urgency, refund. OpenRouter's worked example for Jev.
447 input tokens69 answer tokens
- Decision model$18.77
- Claude Haiku 5.5, answer only$79.20×4.2
- Claude Haiku 5.5 (Low), with thinking$229.20×12.2
Both sides are charged for the same input, which favours the LLM: it also needs its instructions and answer format in the prompt. Only the ticket's token counts are sourced, the other three are my estimates of a typical call.
The last row adds 300 tokens of thinking before the answer. That number is my assumption, not a measurement: Low effort still thinks, and how much depends on the task.
One more caveat: Liquid's docs say d1 bills every question as its own prompt, so for d1 the first row can come out several times higher. Jev's price is per call, which is what the ticket example uses.
Against Haiku the gap is real but modest: about 3 to 5 times with a bare answer, and 5 to 20 times once the LLM thinks. TypeSafe's "one two-hundredth of the cost" is measured against frontier models, and Haiku 5.5 is already a cheap one.
- Decision model (Jev)70 to 500 ms
- Claude Haiku 5.5 (Low), first token7.37 s×15 to ×105
The two bars are not measured the same way. Jev's is the whole answer, end to end, and comes from TypeSafe. Haiku's is only the time until its first token, from Artificial Analysis. The answer is still being written after that, so the real gap is wider than the bars show.
The bigger gap is in time. The cost of a million calls is a line in a budget, but a 7 second wait before the first token does not fit in a request path, and half a second does.
What a call looks like and where it is used
A call is a text and a few typed questions, and the answer is one probability per option. Pick a use to see how it is asked and what comes back.
Model router
A request comes in and someone has to decide who gets it: the strong and expensive model or the weak and cheap one. A decision model reads the task and answers questions such as how hard it is and whether it needs code or tools. Plain code turns the answers into a choice.
Every request that goes to the cheap model saves tokens, and the router itself costs almost nothing. BenchLM describes the same pattern for Jev: one call with 14 choice questions and 5 yes or no questions about a job, and code builds a shortlist of models from the answers.
TaskWrite a Python script that renames 200 photos by the date in their EXIF data.
- ChoiceHow hard is the task?easy 0.12, medium 0.71, hard 0.17
- NoulDoes it need to run code or use tools?0.83
- NoulDoes it need the strongest model?0.09
Smaller jobs on the same pattern: content moderation, ranking candidates.
One warning from the write-ups: hostile text inside the input can push a decision model toward an allowed but harmful answer. In auto mode that matters, so the permission itself should be enforced in code and the model's answer should be only one input to it.
Decision models against LLMs
Artificial Analysis publishes a Decision Index that scores decision models and ordinary LLMs on the same four tests, 25% each: AA-Omniscience-Decision (knowledge and hallucination), HLE-MultiChoice (reasoning), When2Call (tool calling) and WildGuardTest (safety classification). LLMs are included for comparison and drawn hatched on the charts.
Both screenshots below are my own selections from the 28 models on the index, taken before and after d1-ultra came out.
Before d1-ultra
The best decision model here is Jev 1.13 with 58, in second place. The bars on both sides of it are LLMs: Claude Haiku 5.5 (Low) with 61 above and GPT-6 Luna (Low) with 55 below. After Jev come GLIDE and Microsoft-Decision-1 at 54, and the rest go down to Clef-Omni at 44. d1 is at 50.
So a small LLM on low effort still beat every specialised model, by 3 points. Jev came close, but the decision models argued with price and speed, not with the quality of the answers.
After d1-ultra
d1-ultra scores 65. That is 7 points above Jev 1.13, the best decision model of the first shot. Jev had already passed GPT-6 Luna (Low) at 55, but d1-ultra is the first decision model on these charts to pass Claude Haiku 5.5 (Low) at 61, the top bar of the first shot.
I added more LLMs to this shot on purpose, and they stay ahead: Claude Haiku 5.5 on Max effort at 68, Claude Opus 5.5 at 75 and 79.
The honest reading: decision models closed the gap to small LLMs, not to frontier ones. A frontier LLM still answers better. The case for a decision model is that on a task that is only classification, 65 at $0.042 per million tokens may be enough, and the Opus bar costs and waits far more.
Two caveats. The index went public on 13 October, so every number here is a few days old. And d1-ultra reportedly trades some raw accuracy for fewer hallucinations, which is part of how it reaches 65. The newer Jev 1.14 is not in my selection, and coverage of the d1-ultra release gives it 57.
Limits
- 01
They only choose
No text, no tool calls, no multi-step plans. If the answer is a reply or a piece of code, use an LLM.
- 02
They read literally
Implied context is ignored, so the criteria have to be written out in full.
- 03
Weak at exact maths
Arithmetic and date comparisons are unreliable. Compute those in code.
- 04
Confidence is not correctness
A 0.8 tells you how concentrated the answer is. Set thresholds from your own labelled data.
- 05
Closed
Jev and d1-ultra are hosted only, with no published weights, so you depend on the vendor's price and uptime.