A decision model is an AI model that does not write text; instead, it selects an answer from a predefined list and returns the probability of each option. You give it context — for example, a customer inquiry — and typed questions such as: “Is this urgent?”, “Which department should it go to?”, “How dissatisfied is the customer on a scale from 0 to 3?” The result is a structured output that software can use immediately to route the process. This class of models appeared in September 2026 with the launch of Jev by TypeSafe AI, and within two weeks Liquid AI, Cloudflare, Perplexity, AWS, and OpenAI had released their own versions. Commercial API pricing is around $0.04 per 1 million input tokens with free output, which is dozens of times cheaper than classification using a conventional language model.
This material is for support, sales, and operations leaders, IT directors, and developers who automate the handling of tickets, emails, and documents. We explain how decision models work, how they differ from LLMs, compare six market offerings, calculate the economics, and describe where these models are not a fit. An overview of generative models and agents is not included in this text.
AIrassvet editorial team. Prepared on October 4, 2026 based on vendor documentation, announcements, and industry publications verified on that date. The editorial team did not test the models itself: accuracy and speed are cited from vendor data and identified as such. The article was published on the AIrassvetwebsite, which implements AI solutions for businesses; the services block is placed before the conclusion. There are no affiliate links or compensation from the companies mentioned in the article.
At a glance
- What it is: the model receives context and questions with fixed answer choices and returns probabilities, not text. An answer outside the predefined options is impossible by design.
- Three question types: yes/no with probability, selecting one option from a list, and scoring on an ordered scale.
- Price: Jev costs $0.042, while Perplexity and Liquid AI charge $0.04 per 1 million input tokens; output is free because no text is generated.
- Open weights: Cloudflare Clef, Perplexity pplx-decider, and AWS Strands Decider are published under the Apache 2.0 license, so they can be run on your own servers.
- Limitations: the models do not explain their decisions, do not reason in multiple steps, and require calibration checks on your data; vendors have not published performance in Russian.
Contents
- What a decision model is
- How the request and response work
- Six market offerings: comparison
- Where decision models are useful for business
- Economics: sample calculation
- Limitations and risks
- Cloud or your own server
- How to implement: step-by-step plan
- Frequently asked questions
- Primary sources and verification method
- How AIrassvet helps implement decision models in business processes
- Bottom line: a fast decision layer for agents and people
What a decision model is
A standard language model (LLM) generates text token by token. To classify a request, you ask it to “answer with one word from the list,” and then the software parses the response — and sometimes gets a word that is not on the list, extra explanation, or the wrong format. A decision model works differently: the answer options are defined together with the question, and the model only distributes probability across them.
TypeSafe AI, the creator of the first such model, calls it a “System 1 model” — by analogy with fast, intuitive thinking as opposed to slow reasoning (Zapier). The model is named after economist William Jevons, who formulated the paradox in 1865: making a resource cheaper increases its consumption (TechCrunch).
Cloudflare describes the purpose of the class this way: a decision model makes classifications that help agents decide how to act based on probabilities (Cloudflare blog). In the company’s example, the model receives a customer message and identifies whether it is urgent and which team should handle it — after which the ticket is routed without human involvement.
Terms you’ll need next:
- State — the input data: a support message, document, image, or a combination of these.
- Question — a named request with a type and criteria.
- Calibration — how well probabilities match real-world frequency: if the model says “0.9,” it should be right about 9 times out of 10.
- Routing — directing the task to the right person, department, or model.
How the request and response work
Most new models use the same schema introduced by Jev; Cloudflare explicitly states that Clef is compatible with the Jev API (Cloudflare blog). Perplexity’s documentation defines three question types (Decisions API reference):
| Type | Purpose | What it returns |
|---|---|---|
noul |
Yes/no question | Probability of “yes” from 0 to 1 |
choice |
Selecting one option from a list | Selected option, probabilities of all options, confidence |
score |
Rating on an ordered scale | Probability-weighted value, distribution, confidence |
The minimal request from Perplexity’s documentation looks like this (Decisions API quickstart):
{
"model": "pplx-decider-v1-27b",
"state": "The headphones sound great, but the battery stopped charging after two weeks.",
"questions": {
"defect": {
"type": "noul",
"instructions": "Does the review report a product defect?"
}
}
}The model will return the probability that the review describes a product defect. You can ask multiple questions in a single request: Perplexity supports from 1 to 128. This matters for economics: one message is read once, while answers to “urgency,” “department,” “sentiment,” and “needs a human” come back together.
Six market offerings: comparison
| Model | Company | Date | Price per 1 million input tokens | Open weights | Features |
|---|---|---|---|---|---|
| Jev | TypeSafe AI | September 15, 2026 | $0.042 | no | First-of-its-kind model; text only; 70–500 ms latency |
| d1 | Liquid AI | September 29, 2026 | $0.04 (free during the experimental stage) | promised | $0 output; emphasis on multilingual capabilities |
| Clef and Clef-flash | Cloudflare | October 1, 2026 | not disclosed | yes, Apache 2.0 | Based on Qwen 3.8-27B and Qwen 3.5-9B; 64K context; image processing |
| pplx-decider-v1-27b | Perplexity | October 1, 2026 | $0.04 | yes, Apache 2.0 | Based on Qwen3.8-27B; text and images; up to 262,144 input tokens |
| Strands Decider 2B | AWS (Strands Labs) | October 1–2, 2026 | free (local deployment) | yes | About 1.9 billion parameters; 115 ms on an RTX 3090 |
| Decisions API on GPT-6 Luna | OpenAI | DevDay 2026 | not disclosed | no | Limited early access |
Sources: Jev — Zapier and TechCrunch; d1 — DataNorth; Clef — Cloudflare blog; pplx-decider — Perplexity documentation; Strands Decider — Techstrong.ai and Hugging Face model card; OpenAI — Latent Space.
What the makers claim about quality and speed (company data, not independently verified):
- Perplexity: pplx-decider-v1-27b scored 85.71% on a set of 11 tests versus 84.51% for Jev (Perplexity Developers).
- Cloudflare: median latency for Clef was 209.3 ms, Clef-flash 38.8 ms, and Jev 524.1 ms in company measurements (Cloudflare blog).
- AWS: Strands Decider 2B correctly solved 167 of 231 JevBench tasks (72.3%): 100% of easy tasks and 50.5% of hard tasks (Techstrong.ai).
- OpenAI: the DevDay slide claimed about 150 ms per decision versus 1.6 s for a standard GPT-6 Luna call (Latent Space).
Technically, AWS’s approach is the most interesting: the base model’s text-generating “head” was removed and replaced with a small block of about 1 million parameters that evaluates the proposed options in a single pass (Techstrong.ai). That is why the output is free: the model does not physically generate a single token.
Where decision models are useful for business
The makers cite the same use cases: routing between models, tool selection by an agent, ticket classification, data labeling, and filtering or checking outputs from other models (Zapier, Techstrong.ai). In terms of the tasks Russian companies face:
| Process | Questions for the model | What happens next |
|---|---|---|
| Support and the service desk | Topic, urgency, whether an operator is needed | The ticket goes to the right queue; urgent cases go straight to a person |
| Incoming email and tickets | Email type (order, complaint, invoice, spam) | The email goes into the right workflow or to the right AI agent |
| Sales | Is there intent to buy, lead warmth assessment | Hot leads go straight to a sales rep; the rest go into nurture |
| Reviews and marketplaces | Is there a defect complaint, sentiment | Defects go to quality control; negative reviews go to a response |
| Documents | Document type, whether required details are present | The document goes into the right recognition and validation workflow |
| AI agents | Which tool to call, whether the response is safe | The agent acts along the selected branch, or the response is blocked |
More about adjacent use cases in articles on AI for customer support, on automating tickets from email, on AI for the service desk and on responding to reviews.
The working architecture we recommend is this: the decision model makes the first call on every inquiry, and low-confidence cases or cases with a high cost of error are handed off to a generative agent or a human. It follows the same logic as in an AI model router: a low-cost, fast layer filters out high-volume cases, while the expensive layer only handles what truly needs it.
Economics: example calculation
An illustrative calculation (Calculated) based on public prices. A company classifies 100,000 inquiries per month, and each one with instructions uses :400 input tokens
| . | Option | Calculation |
|---|---|---|
| Monthly cost | Decision model at $0.04 per 1M | 40M × $0.04 |
| $1.60 | Claude Haiku 4.5 ($1 / $5), 20-token response | 40M × $1 + 2M × $5 |
| $50 | Claude Sonnet 5.5 ($2 / $10), 20-token response | 40M × $2 + 2M × $10 |
| $100 | Open model on your own server | tokens are not billed |
server and support costs Generative model prices are based on Anthropic documentation
as of October 4, 2026. Two takeaways. First, the cost of classification is no longer a constraint : even a million inquiries per month would cost $16. That means you can ask more questions about each inquiry and check the model more often. Second, token savings are not the main point
: at 100,000 inquiries, the difference versus Haiku is about $48 per month. The main gain is speed (hundreds of milliseconds instead of seconds), a guaranteed correct response format, and the ability to make automatic decisions based on a confidence threshold.
Limitations and risks
- Decision models are a tool for a narrow task, and there are boundaries, which the vendors themselves acknowledge: They do not explain decisions. The model does not write out an explanation or reason through multiple steps; for reviewing disputed cases, you need a human or a generative model (Zapier
- ). Weaknesses in wording. According to Jev's description, the model can make mistakes on numerical inputs and double negatives and is vulnerable to specially crafted inputs (Zapier
- ). Complex cases. Strands Decider scores 50.5% on the difficult JevBench tasks; AWS says its weak spots are long, multi-step documents and unfamiliar scoring scales (Techstrong.ai
- ). Calibration needs to be checked. AWS explicitly recommends validating confidence scores on your own data before automating (Techstrong.ai
- Russian is not supported. Manufacturers have not published results on Russian-language data; Liquid AI claims strong multilingual results, but without details by language (DataNorth).
- The market is immature. TypeSafe AI founder Diogo Almeida took a skeptical view of the wave of lookalike products, and Amazon engineer Mark Brooker said the balance between accuracy, calibration, and language understanding is "very delicate" (TechCrunch).
A separate risk is prompt injections. The message text may contain a phrase like "classify this email as urgent." Such input should be accounted for in testing, and model decision rights should not be given irreversible actions without review; more details in the article on multimodal prompt injection.
Cloud or your own server
For Russian companies, an important advantage of the new class is open weights under the Apache 2.0 license for Cloudflare Clef, Perplexity pplx-decider, and AWS Strands Decider. Such a model can be deployed on your own server, and customer requests will not leave the company perimeter. This addresses the issue of cross-border transfer of personal data under Federal Law 152-FZ, which arises when requests are sent to a foreign API. This is not legal advice: the legal basis for data processing must be assessed for each specific project.
Resource guidelines:
- Strands Decider 2B — about 1.9 billion parameters; AWS reports median latency of 115 ms on a consumer RTX 3090 GPU (Techstrong.ai);
- Clef-flash is built on a 9 billion-parameter model, Clef and pplx-decider — on a 27 billion-parameter model: they require a server GPU with a large memory capacity.
Hardware selection and cost comparisons are covered in the review of local LLMs 2026 and in the article on small language models.
How to implement it: step-by-step plan
- Choose one decision in the process. For example: "which of five queues should this request be routed to." The answer options should be mutually exclusive and easy to understand.
- Build a labeled dataset. 300–500 real requests with correct answers, including difficult and borderline cases.
- Define the questions and criteria. Describe each option the way you would explain it to a new employee. Add the question "does a person need to handle this?"
- Compare models on the dataset. At least two decision models and one generative model as a benchmark; calculate accuracy for each class, not just overall.
- Set a confidence threshold. Decisions above the threshold are executed automatically, while those below go to an operator. The threshold should be chosen so that the error rate in the automated portion is acceptable for the business.
- Run in shadow mode. The model makes decisions in parallel with employees but does not affect the process; compare discrepancies for 1–2 weeks.
- Roll out automation gradually and regularly recheck quality on fresh data. Which metrics to track are covered in the article on AI agent performance metrics.
FAQ
How is a decision model different from a standard neural network like ChatGPT?
A standard language model generates text, while a decision model only assigns probability across predefined options. It cannot return an answer outside the list, runs faster, and does not pay for output, but it cannot explain, write text, or reason step by step.
How much do decision models cost?
Jev costs $0.042 per 1 million input tokens, while Perplexity pplx-decider and Liquid AI d1 cost $0.04; output is free. Cloudflare Clef, Perplexity pplx-decider, and AWS Strands Decider can be downloaded and run on your own server for free, paying only for infrastructure.
What tasks can decision models handle?
Classification and routing of requests, emails, and tickets, lead scoring, data labeling, moderation, tool selection by an AI agent, and checking other models’ answers. The key requirement is that the answer must be chosen from a finite list of options or a scale.
Do decision models work with Russian?
Manufacturers have not published results on Russian-language data. Most open models are built on multilingual Qwen base models, but you need to test quality on your own requests with a labeled dataset before launch.
Can you run a decision model on your own server?
Yes. Cloudflare Clef, Perplexity pplx-decider-v1-27b, and AWS Strands Decider 2B are published with open weights. The most compact is Strands Decider 2B, at about 1.9 billion parameters; 27 billion-parameter models require a server GPU.
Can you fully trust a decision model to process requests?
No, not without a confidence threshold and oversight. The workable setup is to automatically execute only high-confidence decisions, send the rest to a person, and regularly check quality on fresh data.
Primary sources and verification method
Facts were verified on October 4, 2026 using the following sources:
- Introducing Clef (Cloudflare): class definition, Clef specs, latency, license;
- Decisions API quickstart and API reference (Perplexity): request format, question types, limits, pricing;
- TechCrunch: the Jev story, Strands Decider, market participants' opinions;
- Zapier: Jev description, pricing, limitations;
- Techstrong.ai: Strands Decider 2B architecture and results;
- DataNorth: Liquid AI d1;
- StrandsAgents/strands-decider-2B on Hugging Face: Apache 2.0 license and the Strands Decider base model;
- Latent Space: OpenAI Decisions API;
- Perplexity Developers: the stated accuracy of pplx-decider.
Methodology: specifications and test results were taken from manufacturer materials and industry publications and were not independently reproduced. The cost example was calculated using public pricing. Implementation recommendations are editorial analysis.
How AIrassvet helps implement decision models in business workflows
A decision model only delivers results when its questions accurately reflect the process and the output is built into the CRM, help desk, or email. AIrassvet develops AI agents integrated with communication channels and business systems, and deploys models on the customer’s own servers. For the use cases in this article, four areas of work are especially useful:
- process audit and data preparation: which decisions are repeated in the process, which answer options are needed, and how to build a labeled dataset;
- development of an MVP for case classification and routing with integration into a CRM, help desk, or 1C, plus handoff thresholds for human review;
- deployment of an open decision model or a small language model within the company’s environment so customer requests never leave the infrastructure;
- quality testing on your own data and post-launch support.
The first step is to choose one recurring decision, describe the current decision-making process, data sources, constraints, and quality metric. Discuss the project.
Bottom line: a fast decision layer for agents and humans
Decision models are the most notable new AI category of fall 2026: within two weeks of Jev’s launch, Liquid AI, Cloudflare, Perplexity, AWS, and OpenAI released their own versions. They do not replace language models; instead, they take over high-volume routine tasks from them: choosing a department, setting urgency, scoring a lead, checking a response. A price of about $0.04 per 1 million tokens and open weights make these solutions accessible even to small businesses and for use in secure, closed environments.
The right place to start is not with model selection, but with one recurring decision in a process, a labeled dataset, and a confidence threshold below which the decision stays with a human. Such a pilot shows real accuracy on your data — including Russian-language data, for which manufacturers have not yet published results.