An LLM such as Claude or GPT-6 writes text. A decision model such as Jev picks from options you define and returns a probability for each one. If you need to route requests, classify them or approve an AI agent's next step, a decision model is usually faster and gives you more control. Replies, summaries and multi-step work stay with the LLM. Price rarely settles the question: at 100,000 requests a month, every option in our calculation costs less than US$15 in model fees.
Since TypeSafe introduced Jev in mid-September, four more offerings have appeared within four weeks: Cloudflare Clef, Amazon Strands Decider, the OpenAI Decisions API and Microsoft-Decision-1. This comparison is based on vendor information as of 12 October 2026. LindenTech has not benchmarked the models itself; figures from vendor tests are labelled as such.
Jev vs. LLM at a glance
| Question | Decision model (Jev, Clef, Decision-1) | LLM (Claude, GPT-6) |
|---|---|---|
| Output | One of your predefined options, with probabilities | Free text, or JSON following your own schema |
| How it works | One pass over the input; the allowed answers are scored | The answer is generated token by token |
| Confidence | Part of the response, depending on question type | Not by default; has to be derived |
| Billing | Input tokens only | Input and output tokens |
| Good for | Routing, classification, policy checks, approving the next agent step | Writing replies, summarising, extracting fields, calling tools with arguments |
| Not meant for | Generating text, open questions, multi-step planning | Many small decisions where latency and consistent answers matter |
What a decision model does differently
A decision model receives a state, such as a customer email, plus typed questions. Jev has three question types: Choice picks an option from a list, Score rates against a scale and Noul estimates the probability that a statement is true. The Jev documentation explains that all questions in a request are evaluated in parallel and independently against the same state.
A request in Jev format looks like this. Cloudflare's Clef accepts the same format:
{
"state": "Checkout has been failing for every customer for the last hour.",
"questions": {
"urgent": { "type": "noul", "instructions": "Is this support request urgent?" },
"team": {
"type": "choice",
"instructions": "Which team should handle this request?",
"criteria": {
"billing": "Payments, invoices and refunds",
"technical": "Outages, errors and configuration",
"sales": "Plans and upgrades"
}
}
}
}
Under the hood, the model does not generate text. For Clef, Cloudflare describes a single pass over the input, after which the valid answers are scored in parallel. For Strands Decider, Amazon replaced the part of the language model that predicts words with a small scoring head. So the result is always one of your options, though not necessarily the right one.
A wrong category in the correct format is still a mistake. Our article on Jev explains why this matters and how to read confidence values.
When an LLM is still the better choice
As soon as a person has to read the result, you need text: a reply to the customer, a summary, an explanation. A decision model cannot produce that. The same applies to extracting arbitrary fields from a document and to tool calls with arguments. In its Decisions API documentation, OpenAI explicitly recommends Structured Outputs or function calling for those cases.
An LLM is also the right tool when the answer is not on a list you know in advance: investigating the cause of a bug, adjusting a plan, weighing several sources against each other.
The two often work together: the decision model routes the request and approves steps, while the LLM writes wherever text is needed. In our architecture of an AI agent, this is the routing layer in front of the agent itself.
Five decision models in October 2026
| Model | Price per 1M input tokens | Weights and notes |
|---|---|---|
| Jev 1.13 (TypeSafe) | $0.042 | Closed. The first model of this kind, early access, also via OpenRouter |
| Clef, Clef-flash, Clef-omni (Cloudflare Workers AI) | $0.24 / $0.038 / $0.15 | Open, Apache 2.0. Jev-compatible API, images; Clef-omni also audio and video |
| Microsoft-Decision-1 (Microsoft Foundry, OpenRouter) | $0.042 | Not released. Public preview since 9 October, built on Qwen3.5-9B |
| Decisions API with gpt-6-luna (OpenAI) | $0.10 | Closed. Public beta, images, EU data residency for eligible customers |
| Strands Decider 2B (AWS Strands Labs) | No token fee | Open, Apache 2.0. Runs locally, including on CPU and Apple-silicon Macs |
According to the vendors, only input tokens are billed. Sources: Cloudflare changelog of 9 October, Microsoft's announcement, OpenAI's documentation and the Strands Decider repository.
Since the price cut, hosted Clef-flash accepts at most 24,000 input tokens per request. For longer inputs, Cloudflare points to Clef with 64,000 tokens.
Every vendor publishes benchmarks in which its own model leads. Cloudflare puts Clef ahead of Jev; Microsoft puts Decision-1 ahead of everyone. Such tests say little about how well a model routes your requests.
What 100,000 requests a month cost
For the calculation we assume a customer service team that routes 100,000 requests a month. A request is about 250 tokens and the questions with option descriptions about 350, so 600 input tokens in total. An LLM also needs instructions for the response format, which brings it to about 750 input tokens, and returns about 40 tokens of compact JSON. Prompt caching and batch discounts are not included.
| Model | Tokens per month | Cost per month |
|---|---|---|
| Clef-flash | 60M input | $2.28 |
| Jev 1.13 | 60M input | $2.52 |
| Microsoft-Decision-1 | 60M input | $2.52 |
| OpenAI Decisions (gpt-6-luna) | 60M input | $6.00 |
| Claude Haiku 5.5 as an LLM | 75M input, 4M output | $9.50 |
| GPT-6 Luna as an LLM | 75M input, 4M output | $9.50 |
| Clef | 60M input | $14.40 |
LLM prices come from the Anthropic and OpenAI price lists: $0.10 per million input and $0.50 per million output tokens for short prompts in both cases. Each vendor counts tokens with its own tokenizer, so the same text produces slightly different amounts.
So a decision model is not automatically cheaper: in this calculation the large Clef costs more than Claude Haiku 5.5. The gap between the cheapest and the most expensive option is less than $13 a month.
Wrong decisions weigh far more. If 3 in 100 requests land with the wrong team and each correction takes two minutes, that adds up to 100 working hours a month at 100,000 requests. Improving accuracy on your own data is therefore worth more than hunting for the lowest token price.
Why milliseconds matter in agents
An agent rarely makes just one decision. In its Decision-1 announcement, Microsoft does the maths: 100 extra milliseconds across 20 sequential decisions add two seconds of waiting. At 500 milliseconds per step that becomes ten seconds; at 40 milliseconds, less than one.
Vendor figures vary widely. Cloudflare reports a median of 38.8 ms for Clef-flash and around 150 ms for Clef at about 800 input tokens after the 9 October optimisation. OpenAI speaks of answers about ten times faster than through the Responses API, without stating test conditions. On 12 October, OpenRouter showed a median of 0.20 seconds for Jev and 0.35 seconds for Decision-1, network included. Measure latency from your own region and infrastructure.
Where the decision is computed
For companies in the EU, where the data is processed often matters more than the model. The vendors handle this differently:
- Strands Decider and Clef are released under Apache 2.0 and can be run on your own infrastructure, so the data stays with you. According to its documentation, Strands Decider 2B also runs without a graphics card, while Clef needs considerably more compute.
- For the Decisions API, OpenAI offers Zero Data Retention and data residency in the EU and Switzerland for eligible customers. Cloudflare commits to not reading, storing or training on requests to Clef.
- For Jev and Decision-1, confirm the processing location, retention and data processing agreement with the vendor before use.
We have collected the questions to ask in our guide to running AI in the EU.
Decision model, LLM or both?
- Is the right answer on a fixed list? Then a decision model fits.
- Will a person read the result? Then you need an LLM or a text template.
- Do many decisions run one after another? Then latency matters, and the decision model has the edge.
- Do you need a threshold for "unsure, hand over to a person"? Decision models provide a probability for that. Set the threshold using your own test data.
- May the data leave your company? If not, only self-hosted models with open weights qualify.
If the decision can be written as a fixed rule, such as whether a deadline has passed, you do not need a model at all. More on that in automation without AI.
How to compare Jev, Clef and an LLM fairly
Collect 200 to 300 anonymised requests and record the correct decision for each. Hold back a portion that you do not look at while writing the questions. Ask every candidate the same questions with the same options. Because Clef understands the Jev API, Cloudflare says switching between the two only means changing the model ID.
Compare four values:
- accuracy per category, not just overall;
- the share of decisions that are wrong with high confidence;
- 95th-percentile latency from your region;
- cost per correctly routed request, including correction effort.
Use a simple rule or your current solution as the baseline. Only when a model beats that baseline on your data is it worth introducing.
Frequently asked questions
Is Jev an LLM?
No. Jev is built on language-model technology but does not generate text. It scores predefined answers and returns probabilities.
Can I replace Jev with Clef?
Cloudflare states that Clef is compatible with the Jev API. Technically the switch is simple. Whether the decisions are as good is something only a test with your data can show.
Which decision model runs locally?
Amazon's Strands Decider 2B and Cloudflare's Clef are available to download with open weights. Strands Decider is small enough for an ordinary server or a Mac.
Does a decision model replace an AI chatbot?
No. It can decide where a request belongs and whether a reply is approved. The reply itself is still written by a person, a template or an LLM.
Is a decision model always cheaper than an LLM?
Not necessarily. In our calculation Clef costs more than Claude Haiku 5.5. For most companies, model fees are small compared with the effort of fixing wrong decisions.
If you want to check whether a decision model fits your customer service or AI agent, talk to us about AI automation with LindenTech. Bring one recurring decision: what data arrives, which answers are allowed and what a mistake costs.