Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarHabr AI→ original

Decision models answer in milliseconds: Jev costs $0.042 per million tokens

Decision Models are models for making decisions: they don't write an answer in words, they return probabilities over options right away. For example, given a complaint like "the money was debited, but there's no order," the model replies "tech support — 87%, accounting — 10%" and that's all. The first of them, Jev from TypeSafe AI, cost $0.042 per million input tokens at launch, and output tokens weren't billed. That makes it cheaper to hand routine choices in agents to such a model and reserve complex cases for a large LLM.

AI-processed from Habr AI; edited by Hamidun News
Decision models answer in milliseconds: Jev costs $0.042 per million tokens
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Three task typeschoosing an option, yes/no, and scoring on a scale
Key metric #2Answers in tens to hundreds of milliseconds: the network doesn't generate text
Key metric #3Gemini Decision 1 from Google costs $0.10 per million input tokens

Imagine you're building an AI agent for customer support. A customer writes: "I paid for my order yesterday, the money was debited, but the order isn't in my account." The agent only needs to pick who should get the ticket, yet it runs an expensive language model for this, which first reasons through three paragraphs.

In autumn 2026, a separate class of solutions appeared for this problem: Decision Models. They don't talk to people and don't write text; they return probabilities over predefined options. The first one, Jev from TypeSafe AI, launched in September, and it costs $0.

042 per million input tokens (a token is roughly a piece of a word), while output data wasn't billed at launch. For comparison, Gemini Decision 1 from Google charges $0.10 per million input tokens.

What such a model does in a simple example

The name refers to Daniel Kahneman's book: "System 1" is fast, intuitive thinking, and "System 2" is slow and deliberate. Decision Models, or System One Models, are the "fast" part of AI. For a complaint about debited money, the model returns only numbers: tech support — 87%, accounting — 10%, delivery — 2%, sales — 1%. No explanations, just something a program can use immediately.

Most models solve three types of tasks:

  • Choice — pick one option from a list: who should receive the ticket or which action to take next.
  • Yes/No — check a condition: does a document meet the requirements, is there an error in the answer, does a human need to get involved.
  • Score — rate on a scale: how urgent a task is or how well a document matches a query.

Many models accept several questions about one object at once and answer them in a single request.

How this differs from a classifier and from an LLM

Classifiers have existed for a long time: logistic regression, random forest, CatBoost. But an anti-fraud model trained on transactions won't evaluate the quality of a support reply: a new task requires a different model or fine-tuning. A Decision Model tries to combine the versatility of an LLM with the convenience of a classifier. The answer options and criteria are set directly in the request, with no retraining. The author notes that for stable, well-defined tasks, a specialized classifier may turn out to be more accurate, faster, and cheaper.

New models have four advantages over a regular LLM:

  • Speed. An LLM generates its answer token by token, even when it needs a single "yes" or "no." Many Decision Models evaluate the options in one pass through the network, and the answer arrives in tens or hundreds of milliseconds.
  • Cost. Liquid d1-ultra costs roughly the same as Jev. For a single request the difference is small; for millions of decisions per day it becomes noticeable.
  • Format. The answer is limited to a predefined set of options. But a correct format doesn't guarantee a correct decision: these models can still be wrong.
  • Confidence. Probabilities allow simple cases to be handled automatically, while complex ones go to a strong model or a human. If a model says "90%" a hundred times, it should be right in about 90 of those cases. This needs to be verified on your own data.

Who has already launched

The wave began on September 15, 2026, with Jev; TypeSafe AI was founded by former OpenAI employee Diogo Almeida. Then came Tev1 from Together AI (September 22), the open Kev (September 24), Clef from Cloudflare (October 1), the multimodal d1 from Liquid AI (October 5), Liquid d1-ultra (October 15), the Decisions API from OpenAI (October 22), and Gemini Decision 1 (November 11). On November 17, Liquid AI released d1.2.

Available via API: Jev, Liquid d1.1 and d1-ultra, Gemini Decision 1 (text, images, audio, video, and PDF), the OpenAI Decisions API based on GPT-6 Luna, Perplexity Decider and Decider Flash v1.3, as well as the Clef, Clef-flash, and Clef-omni family.

You can run locally: Kev 4B based on Qwen3.5, Liquid d1-2B and d1-Omni-2B, the experimental d1-omni-600M with 600 million parameters, Perplexity Decider Flash, the Laya family based on ModernBERT and mmBERT, open weights for Clef and Clef-flash, and Tev1 4B and 0.8B. Internally they are built differently: some don't generate tokens at all, while others are fine-tuned language models with very short answers.

What to do with this

The most obvious pattern for agents is "brain and reflexes." A powerful LLM builds the strategy and plans, while a small Decision Model constantly picks the next step and calls the large model only in non-standard situations. Other applications include routing requests between agents, selecting tools, checking the answers of other models, evaluating code, moderation, and detecting attempts to manipulate an agent.

There are also limitations. The author hasn't yet shown their own comparison tables and promises them in a follow-up, so quality figures will need to be verified independently. Start with a small set of labeled examples from your own pipeline and compare the models' choices against manual labeling. Then check calibration: does the stated confidence match the actual share of correct answers? Only after that should you set a threshold below which a decision goes to the large model or to a human.

Story

Key Questions Answered

What such a model does in a simple example?

The name refers to Daniel Kahneman's book: "System 1" is fast, intuitive thinking, and "System 2" is slow and deliberate. Decision Models, or System One Models, are the "fast" part of AI. For a complaint about debited money, the model returns only numbers: tech support — 87%, accounting — 10%, delivery — 2%, sales — 1%. No explanations, just something a program can use immediately.

How this differs from a classifier and from an LLM?

Classifiers have existed for a long time: logistic regression, random forest, CatBoost. But an anti-fraud model trained on transactions won't evaluate the quality of a support reply: a new task requires a different model or fine-tuning. A Decision Model tries to combine the versatility of an LLM with the convenience of a classifier. The answer options and criteria are set directly in the request, with no retraining.

Who has already launched?

The wave began on September 15, 2026, with Jev; TypeSafe AI was founded by former OpenAI employee Diogo Almeida. Then came Tev1 from Together AI (September 22), the open Kev (September 24), Clef from Cloudflare (October 1), the multimodal d1 from Liquid AI (October 5), Liquid d1-ultra (October 15), the Decisions API from OpenAI (October 22), and Gemini Decision 1 (November 11). On November 17, Liquid AI released d1.2.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…