Decision Models Explained: From Generative AI to AI That Makes Decisions
A2Agent Team ยท 2026-11-17T00:00:00Z
Most of what we call AI right now is built on large language models. ChatGPT, Gemini and Claude read text, write text, answer questions and reason through problems. Agents change what we ask of them. An agent handling a customer request has to decide which model should process it, whether it needs a web search, whether the input is risky, and whether the task is ready for the next step. None of those decisions needs a paragraph of text. What the software needs is a clear answer it can act on.
That's the job of a new kind of model: the decision model.
What a decision model is
A decision model takes some input and a set of questions with predefined answers, and returns a probability for each answer. It doesn't write sentences. The output is a choice, a score or a yes/no, with numbers attached, ready for code to use.
Take a customer support platform that receives a complaint. An LLM can read it, understand it and write a summary. If all the platform needs is to send it to the right team, that summary is wasted work. A decision model can return something like billing 0.08, technical 0.85, sales 0.07 and the platform routes it to technical support.
Most decision models handle three kinds of question:
| Question type | What it returns | Example |
|---|---|---|
| Yes/no | A probability from 0 to 1 | Does this reply contain a misleading claim? |
| Multiple choice | A probability for each option | Which model should handle this request? |
| Score | A position on a scale you define | How urgent is this customer complaint? |
LLMs can already do this with careful prompting and JSON output. The difference is that decision models are built only for it. Because the possible answers are fixed before the call, they can't return something outside the list, and because they don't generate text token by token, they're much faster and cheaper.
They also aren't reasoning models or agents. Reasoning models work through hard problems step by step. Agents coordinate models and tools to finish a task. A decision model is a small part inside that system that picks between options quickly. It complements an LLM and doesn't replace it.
How the category took shape
The need isn't new. Machine learning classifiers, recommendation systems and model routers have made these kinds of choices for years. A well-known open-source example is RouteLLM from the LMSYS team, a framework for training routers that send each request to a cheaper or stronger LLM depending on how hard it looks. What changed in the fall of 2026 is that several companies started selling decision-making as its own model and API.
September 15: TypeSafe AI launches Jev
TypeSafe AI, a startup with $40 million in funding led by former OpenAI researcher Diogo Almeida, released Jev in early access on September 15, 2026, and called it the first "System One model." The name refers to Daniel Kahneman's Thinking, Fast and Slow, where System 1 is fast, intuitive thinking and System 2 is slow and deliberate. In TypeSafe's framing, chat LLMs are System 2 tools and Jev is System 1.
You send Jev some state, such as a JSON object or a sentence like "My card was charged twice," along with typed questions. It answers all of them in one pass with calibrated probabilities. TypeSafe says it responds in 70 to 500 milliseconds and costs $0.042 per million input tokens, with output free. In the launch demo, Jev answered in 0.114 seconds where GPT-5.6 Terra took 8.6 seconds. The name is a nod to the economist William Stanley Jevons and the Jevons paradox, the idea that making something cheaper makes people use more of it.
September: open-source decision models
The open-source community moved quickly. The vLLM Semantic Router project released Decision 1.0, a family of open decision models under the Apache 2.0 license. The largest, Decision 1.0 Lux 9B, adds a decision head to Qwen3.5-9B and returns probabilities for choice, yes/no and score questions. Open models mean teams can run decisions on their own hardware, and they tie decision models closely to routing, the job the vLLM project was built for.
October 5: Liquid AI launches d1
Liquid AI released an experimental version of d1 in late September and launched it officially on October 5, now with image input. Like Jev, it reads the input and questions in one forward pass and returns a probability for each answer without generating tokens. Text decisions take 200 to 300 milliseconds. Liquid's examples include sorting defective parts on a production line, flagging support tickets as a SQL function, filing documents into folders, picking a web agent's next action, and trimming tool output from a coding agent's context, which removed 52% of the tokens. It costs about $0.04 per million input tokens. Liquid later added d1-ultra, a proprietary multimodal version that leads Artificial Analysis's Decision Index.
October: Perplexity's Decider
Perplexity released its Decider family for classification, routing, moderation and scoring. Decider V1.2 is a 27B-parameter model built on Qwen3.8-27B. It reads text and images, supports up to 255 answer options, and is available both through Perplexity's API and as open weights on Hugging Face under Apache 2.0. At about $0.02 per million input tokens, it's one of the cheapest options, which matters for decisions that run thousands of times a minute.
November 11: Google launches the Gemini Decisions API
On November 11, 2026, Google launched the Gemini Decisions API, powered by a new model called Gemini Decision 1. It accepts text, images, audio, video and PDFs together, and can answer up to 1,000 independent questions about the same input in one request. It returns probabilities for multiple-choice questions and weighted scores for rubric questions.
Google reports 150 to 250 milliseconds end to end for inputs up to 16,000 tokens, and the time barely changes between 1 and 100 questions. It costs $0.10 per million input tokens, with output free. Google's own example: running 20 checks on a 4,000-token record costs about $0.50 per 1,000 records in one call, against about $8.10 if each check were sent as a separate LLM request. Google already uses it inside Gemini Notebook to route requests between models and reasoning levels, which it says cut latency on routine queries by up to 40%.
OpenAI, Amazon and Microsoft have also shipped decision products. OpenAI's Decisions API runs on GPT-6 Luna and accepts up to 400 questions per request. Amazon released Strands Decider 2B as an open model, and Microsoft has Microsoft-Decision-1.
| Company or project | Model | What stands out |
|---|---|---|
| TypeSafe AI | Jev | Started the "System One" category; $0.042 per 1M input tokens, output free |
| vLLM Semantic Router | Decision 1.0 | Open models under Apache 2.0, built for routing |
| Liquid AI | d1, d1-ultra | Text and image decisions; d1-ultra leads the Artificial Analysis Decision Index |
| Perplexity | Decider V1.2 | 27B, open weights, about $0.02 per 1M input tokens |
| Gemini Decision 1 | Text, image, audio, video and PDF; up to 1,000 questions per call | |
| OpenAI | Decisions API (GPT-6 Luna) | Up to 400 questions per request |
| LMSYS | RouteLLM | Open-source routing framework, an earlier step toward the same goal |
RouteLLM is a different kind of thing from Jev or Gemini Decision 1. It's in the table because it shows that automated model selection had demand well before anyone sold a dedicated decision model.
What decision models are good for
Decision models won't replace the jobs LLMs do. They're for large numbers of repetitive, well-defined judgments that have to be fast and consistent, which is common in agents, enterprise software and AI infrastructure.
Model routing
Routing is the most obvious use. Models differ in reasoning, coding, context length, speed and price. An API gateway that connects to many models has to decide, for every request, whether it's a simple question, a coding task or a hard reasoning problem, and send it to the right model.
The old way was to ask an LLM to analyze the request, write an explanation and name a model, then parse the answer. A decision model takes the request and a list of models and returns a probability for each, in a fraction of a second, for a fraction of a cent. For API gateways and model aggregators, this is the most practical use of the technology today.
Classification and moderation
Companies receive huge amounts of unstructured input: customer messages, reviews, documents and images. Before any of it can enter a workflow, it has to be sorted. An e-commerce site with thousands of support requests a day needs to know which ones are about product quality, which about delivery and which about refunds, and which are urgent or abusive. Decision models handle this kind of sorting well, along with sentiment and urgency scoring, product defect detection from images, policy screening and document filing.
Agent workflows
Agents now run multi-step tasks: searching, querying databases, calling APIs and switching between models. At every step, the system has to decide what to do next. A research agent, for example, has to decide after each search whether it has enough evidence, whether the results meet the bar, and whether a conflict calls for a stronger model to check.
A decision model can make those calls: whether to search again, whether results are good enough, whether to bring in another model, whether to ask the user for more information, and whether the task can finish automatically or needs a person. The LLM still handles understanding the goal, planning and writing. Splitting the work this way makes agents more modular, with each model doing what it's best at.
Risk checks and quality control
As AI systems act on their own, wrong decisions cost more. A payments platform has to spot suspicious transactions. A company has to check that automated replies follow its policies. An agent has to judge whether an action is too risky. Decision models can return scores for these checks and pass them to the rest of the system.
A common pattern is a confidence threshold. When the model is confident, the system proceeds. When it isn't, the request goes to a stronger LLM or a person. A confidence score isn't a guarantee, though. Decision models still make mistakes, especially when the criteria are vague, the input is incomplete or the task needs specialist knowledge. TypeSafe calls Jev "hallucination-free," but as The Register pointed out, that's only because it can't produce free text. A wrong probability is still a wrong answer. Test on your own data, and match the amount of human review to the cost of a mistake.
We use this pattern ourselves. InkTrace, our Chrome extension for spotting AI-written articles, runs on Jev.
Who should care
Decision models aren't for people who want to chat with an AI. They're for teams building AI into software:
- Agent developers, who need a fast way to choose between models, tools and actions without generating text at every step
- API gateways and model aggregators, which have to balance capability, cost, speed and reliability across many upstream models on every request
- Businesses with large volumes of repetitive data, in e-commerce, finance, support, content platforms and SaaS
- Product managers and automation designers, who can split a workflow into understanding, generation, evaluation, decision and execution, and pick a different model for each stage
They're a poor fit for writing articles, open-ended research, explaining complex ideas or long conversations, which still belong to LLMs. Decision models work best when the goal is clear, the possible answers can be listed in advance, and the result feeds directly into the next step of a program.
What comes next
AI is moving from answering questions to finishing tasks, and fast, reliable decisions now matter as much as good writing. Future systems will probably combine several kinds of model: LLMs for language and content, reasoning models for hard problems, and decision models for choosing options, tools and next steps. Within two months of Jev's launch, Liquid AI, Perplexity and Google had shipped their own decision models, OpenAI, Amazon and Microsoft now have decision products too, and there are open-weight options for teams that want to run their own.
The limits are real. Decision models need well-defined criteria, their probabilities aren't guarantees, and they struggle with ambiguous or open-ended work. They'll sit alongside LLMs, not take their place. The next round of competition may depend less on who builds the single strongest model and more on who combines different kinds of models well.
Build with multiple models on A2Agent
A2Agent gives you DeepSeek, GLM, Kimi, MiniMax, Qwen and other models through one API key, with OpenAI-compatible and Anthropic-compatible endpoints, so you can put a fast model on routine steps and a stronger one on hard reasoning. See the Models page for current availability and prices. New accounts get $5 in free credit.
Sources
- TypeSafe AI debuts model for machines that plays Doom (The Register, September 16, 2026)
- What is Jev? TypeSafe AI's System One model (DigitalOcean)
- Decision 1.0 Lux 9B (vLLM Semantic Router, via Made by Agents)
- Introducing d1: the most capable decision model, now with vision (Liquid AI, October 5, 2026)
- Perplexity pplx-decider-v1.2: is it worth using? (MindStudio)
- Gemini Decisions API: fast, calibrated multimodal decisions (Google)
- Google launches Gemini Decision 1 API for 1,000 questions per call (Dataconomy, November 12, 2026)
- Top 10 AI decision models in 2026 (DataCamp)
- RouteLLM (LMSYS)