Decision models, explained

Decision models in brief

How it works

A language model answers by generating tokens one at a time, each conditioned on the last. Most of the cost and latency of an LLM call goes into that loop. Yet many steps in a real application never needed prose. Which of five queues gets this ticket? Is this message spam? May the agent run this refund? A decision model is a deep learning model built for that shape of problem. The industry also calls them "System One" models, after the psychology term for fast, intuitive judgment. They are the counterpart to the slow, deliberate System Two thinking that reasoning models spend tokens on.

The request: state plus typed questions

A call has two parts. The state is the thing being judged: a support ticket, a chat transcript, a JSON record or, on newer models, an image. The questions say what to decide about it. The vendors' question types are converging on a small set:

• Choice: pick one option from a list you supply, with a probability for every option.
• Score: place the input on an ordered rubric you define (Jev allows up to 10 levels), returning the probability of each level.
• Yes/no: the probability that a statement about the input is true.
• Box: on Jev, the bounding box of a target in an image.

The model answers every question in a single forward pass instead of a token-by-token loop. That is where the speed comes from. Vendors claim up to roughly ten times faster than a reasoning model, and TypeSafe quotes 70 to 500 milliseconds end to end for Jev. The output is only a few dozen tokens of structured data, so the vendors bill input only.

Calibration and thresholds

The probability is half the product. A model is calibrated when its stated confidence matches its hit rate: across many answers given at 0.8, about 80% should be right. Calibration holds on average, not per answer, so any single answer can still be wrong. It is what lets you set a threshold. Act automatically above it, and send everything below it to a bigger model or a person. Some vendors also train their models to abstain. Base Labs, for example, included training examples that teach the model when to say "I don't know." Calibration varies a lot between products, and none of the vendors' figures has been reproduced independently yet. Pick your threshold from your own labeled data, not from a launch post.

What they cannot do

A decision model returns answers to the questions you asked and nothing else. It writes no prose, JSON bodies or explanations, shows no reasoning, makes no tool calls, and cannot hold a conversation or plan a multi-step task. Jev's documentation also warns against exact arithmetic, date math and numeric threshold comparisons; do those in code. If a step needs any of that, it belongs to an LLM, and the decision model's job is to decide whether that LLM is needed at all.

Two ways to build one

The products so far take two routes. Some are trained for the job: Base Labs post-trained open Qwen3.5 base models on 6.21 million decision examples to make mammoth-1 and elephant-1. OpenAI took the other route. Rather than train a dedicated model, it reconfigured an existing ChatGPT model, Luna, for fast, cheap inference behind its Decisions API. Purpose-trained models currently lead the public leaderboards. The adapted model brings OpenAI's platform, billing and data controls with it.

The cost math

Pricing is per input token, so the shape of your prompt drives the bill. At Jev's list price, a 447-token support ticket with three questions costs about $0.000019, or roughly $19 per million tickets. Two details catch people out. First, some vendors charge the full input again for each question (Liquid does, images included). Second, a long policy or rubric sent with every call is billed every time. On prompt-heavy workloads, that can erase the savings over an ordinary LLM call that uses prompt caching. Measure on real traffic before you commit.

There is no agreed definition of a good decision model yet. Every product trades off four things: intelligence, cost, calibration and speed. The right balance depends on the workload, which is why the public rankings disagree with each other. As with any benchmark, use leaderboard position to build a shortlist, then decide with your own evals on labeled examples.

Where it sits in the AI stack

A decision model sits at the branch points of an application or agent. It chooses a path and escalates only when it is unsure:

The common jobs so far:

• Routing and triage: send each request to the right queue, model or workflow before anything expensive runs.
• Classification and tagging at scale: label millions of records where an LLM call per record would cost too much.
• Gating agent actions: approve or block a tool call before it executes, as a fast layer of guardrails.
• Checking LLM output: score a draft against a policy or rubric. This is a cheap first pass before (or instead of) LLM-as-judge.
• Real-time and voice: make turn-by-turn choices inside a conversation, where an extra second of LLM latency is audible. Early OpenAI customers such as Speak reported large latency cuts.

Key tools and implementations

• Jev (TypeSafe)

  The model that created the category. Early access opened in September 2026, and Jev 1.14 (October) added image inputs and "smart-if" statements. It is the most-queried decision model on OpenRouter and tops JevMan. TypeSafe was valued at $7.6 billion in October 2026.

• OpenAI Decisions API

  Announced at OpenAI's developer day on September 29 and opened to all customers on October 22, 2026. It runs on a reconfigured ChatGPT Luna rather than a dedicated model. It ranks below Jev on both public indexes, but brings OpenAI's platform and billing with it.

• d1-ultra (Liquid AI)

  Released October 15, 2026. It takes text and images in 44 languages. Liquid reports that it beats Jev 1.14 in all eight areas of its own test suite, with lower calibration error. It is served through Liquid's API, with OpenRouter and Vercel listings announced.

• mammoth-1 and elephant-1 (Base Labs)

  Baseten's research team released these 27B and 9B models on October 21, 2026, with open weights on Hugging Face. Base Labs reports mammoth-1 first on the Jev Decision Index, and the small elephant-1 level with d1-ultra on the Artificial Analysis index.

• Quality indexes

  The Artificial Analysis Decision Index, the AIM System One Index and the Jev Decision Index (over 200 models) each rank decision models with their own test mix. Expect the order to differ from one index to the next.

• JevMan and usage boards

  JevMan has models play Pac-Man, which tests high-stakes choices made within a few hundred milliseconds. OpenRouter's decisions leaderboard ranks models by usage, which shows popularity, not quality.

At a glance (October 2026)

Prices and scores are the vendors' own or come from early press coverage, and they move fast. Check the vendor's page before you build on them.

Related entries

• Reasoning models Models trained to work through problems step by step before answering, trading time and tokens for accuracy.
• Model benchmarks Standardized tests that score model capabilities, useful for rough comparison but easy to overfit and game.
• Choosing a model A framework for weighing capability, latency, cost, privacy, and openness when picking a model for a task.
• GPT OpenAI's family of closed-weight language models, served through its API, Microsoft Foundry, Amazon Bedrock, and the ChatGPT products.
• Tool use (function calling) Letting a language model request actions by emitting structured calls that your code executes and whose results feed back in.