AI decision models: Jev, Clef, uses and costs

Sorting an email, judging whether an agent’s action is risky, choosing which department will handle a request: none of these decisions needs a written sentence. That is the bet behind AI decision models, a category opened on 15 September 2026 by Jev, from the American start-up TypeSafe AI. According to Vercel, Jev became the most quickly adopted model on its AI gateway in a single day, and OpenAI, Cloudflare and Amazon presented their own versions about two weeks later.

Here is what a decision model is, what the October 2026 offerings are worth, what one is used for in your line of work and what we think of it. Then the implementation, the way we carry it out: designing the questions, choosing the infrastructure, the cost, and monitoring once the model is in service.

What is an AI decision model?

Cloudflare defines it as a model that makes classifications to help an agent decide how to act, based on probabilities. The exchange takes four steps:

The example in the Clef announcement comes down to three questions about a single message: “Checkout has been failing for every customer for the last hour.” Is the request urgent? Which team should handle it: billing, technical or sales? How severe is the impact on the customer, from “No impact” to “Critical”? The model returns a probability for each option of each question, and the code files the ticket, raises an alert or hands the case to a human.

Two differences matter. With a traditional classifier, adding a category meant retraining the model; a decision model receives its categories in the request, Cloudflare points out. And unlike a large language model, it writes nothing: Clef reads the input in a single pass, then scores all the permitted answers in parallel, without producing text word by word, which is why it is fast. TypeSafe places Jev among its “System One models”, a reference to the psychologist Daniel Kahneman’s System 1: fast, intuitive thinking, whereas the large models that “reason” play the part of System 2, slow and deliberate.

Are decision models deterministic?

As to the form of the answer, yes. Cloudflare contrasts these models with large language models, which it calls “largely non-deterministic”, and its first demonstration exposed “deterministic probabilities”. A decision model never strays from the options it is given, and TypeSafe says it designed its own to return answers that are stable from one evaluation to the next, in its design guide. With the same state and the same questions, the answer is supposed to stay the same, which makes it possible to test it and to cache it.

As to the judgement, no. TypeSafe’s promise that Jev “can’t hallucinate” means that Jev never answers outside the list; nothing stops it from picking the wrong option in it, The Register points out. The answer remains a probability, and the whole point is that it can be trusted: a well-calibrated model that says 90% should be wrong about one time in ten. Cloudflare trained Clef with a Brier loss function to refine that calibration. On the Strands Decider model card, Amazon warns that its calibration was fitted on short classification tasks: measure on your own traffic before you trust a threshold.

Deterministic in form, probabilistic in substance: treat a decision model as a very fast sorter whose error rate you measure before you hand it the sorting.

Jev, Clef, Strands, OpenAI: which decision models are available in October 2026?

TypeSafe AI came out of stealth with Jev and $40 million in funding, then raised $870 million on 9 October, at a valuation of $7.5 billion, from Andreessen Horowitz and Sequoia Capital, SiliconANGLE reports. A third of the Fortune 500 companies are said to be using Jev already, according to the vendor. Less than a month after the launch, the market looks like this:

The prices come from the price lists of Jev, Workers AI, OpenAI and Perplexity, consulted on 11 October 2026; none of them charges for output tokens. Clef’s weights are published on Hugging Face, as are those of Clef-omni, Strands Decider and Perplexity’s model. As for the Decisions API, Sam Altman presented it on 29 September, at the conference where OpenAI also launched its $500 ChatGPT plan. He sees it as a way to focus the Luna model on a single choice and “make it extremely fast” while keeping image understanding, broad language support and safety protections, TechCrunch reports. The API has since moved to public beta.

On its own measurements, Cloudflare announced at launch a median latency of 209 milliseconds for Clef, 39 for Clef-flash and 524 for Jev, then made Clef 1.7 to 2 times faster on 9 October, according to its second announcement. On TypeSafe’s four business evaluations, Clef beats Jev on three of them and loses on agent trace observability:

These are gaps of a few points, measured by Cloudflare itself and not yet reproduced for the official leaderboard, The Register notes. Clef-flash, which is cheaper than Jev, is one point ahead of it on customer service and level with it on security, but loses nearly five points on invoices (57.1) and nearly two on agent traces. On the public benchmarks in the same table, the gaps are sometimes sharper, in both directions: Clef wins by a wide margin on BANKING77 (94.2 against 79.7), Jev on When2Call (81.0 against 72.4). The table already lists other models of the kind, such as Kev 9B or Laya: the market, Cloudflare writes, is “getting increasingly saturated”.

Cloudflare also gives an in-house example. Its threat intelligence team has Clef classify domains: Clef opens the page with its Browser Run browser and answers, for example, “fashion website” at 95%, “ecommerce” at 85% and “phishing” at under 1%. All in 2.2 seconds, against 4.7 for gpt-oss-120b, its fastest general-purpose model, which returned only two categories. Added together, these probabilities exceed 100%: each category is judged separately, and a site can be both a fashion website and an online shop.

The head of TypeSafe, Diogo Almeida, does not yet see a real competitor. An OpenAI alumnus who co-authored the 2022 paper on InstructGPT, the training method that came before ChatGPT, he finds that these copies look more like engineers “wanting to implement a cool architecture” than a team devoted to making intelligence useful, he told TechCrunch.

What is a decision model used for, by line of work?

A decision model is useful wherever the same decision comes up hundreds or thousands of times, on freely written text, and determines what happens next. TypeSafe’s use-case map sorts them by sector, often with an exit to a human. Here is how that reads for companies:

The first published cases point the same way. At Vercel, a classifier that checks the safety of commands switched from an OpenAI model to Jev, with answers 5 to 18 times faster, and more accurate, according to TechCrunch. At Jack & Jill, a recruitment platform where agents match candidates with companies, Jev replaced Gemini 3.1 Flash Lite to shortlist candidates: costs cut eightfold and sorting twice as fast, at comparable quality, according to the case study published by TypeSafe. At Cloudflare, Clef already closes unwanted tickets on the public documentation repository, spots phishing among the extensions of its content management system and looks for personal data, such as ID numbers, according to the 9 October announcement; the company also wants to fine-tune it to sort its support requests, to assess the reports handled by its trust and safety team, or to tell good bots from bad ones. And the OpenRouter gateway offers Jev Router, which chooses the model and the reasoning effort for each request. Clef also reads images, and Clef-omni sound: a product photo, a screenshot or the noise of a machine can be part of the state.

For a small business, the first question is volume. A decision made twenty times a day can often be handled with a written rule, or by hand. A decision model becomes interesting when the volume exceeds what one person can sort without spending their days on it, or when the answer has to come back in a fraction of a second, in an interface or in an agent’s loop.

What are the promises of decision models worth?

For their supporters, these models fix a mismatch: asking a model trained to converse for a “yes” means paying for prose to get a Boolean. “We have lightning in a bottle, and yet it is not useful,” Diogo Almeida sums up in TechCrunch, because large models are optimised for human language. The second argument is probability. TypeSafe puts it this way: “If a model can do a task 95% of the time but doesn’t say when it’s in the 5%, it can’t automate that task.” The CTO of Bryo AI also sees Jev as the only model that returns a genuine probability, even though, in his test on sorting business emails, Gemini proved a little more accurate, at 10 to 20 times the price.

Sceptics start with the figures. Those shown on Jev’s home page come from evaluations built by TypeSafe, which admits as much in its announcement: these gains are probably at the higher end of the real-world range, and some bias cannot be ruled out. The measurements published by users themselves were collected by the team behind the software OpenChamber from 12,759 posts on X, including 215 speed measurements and 180 price measurements; a survey of user reports and not a benchmark, its authors warn.

The gains users report are still considerable, but far from the home page. For Armin Ronacher, CTO of Earendil, Jev “delegates the hallucination problem a little bit to the user”, who has to decide what to do with a 50% answer (TechCrunch).

Nor is the idea new: classifiers have existed for years, and the most widely seen criticism of the launch, according to OpenChamber, reduced Jev to a smarter switch statement. The novelty, if the promises hold: nothing to train, categories defined in the request, the text understanding of a large model, and a probability you can trust. The adoption, for its part, is real: according to Vercel, nearly 13% of the paying teams on its gateway were using Jev after 24 hours, twice as many as the GPT-5.6 family at the same stage, and the next test will be whether that adoption lasts.

What does this change for the AI industry?

First effect: the format became commonplace in two weeks, to the point that Clef reuses Jev’s API, so you can swap one for the other without rewriting your code. Prices followed: Perplexity charges half what Jev does, and on 9 October Cloudflare brought Clef-flash below Jev’s price, by shrinking its context window. The same day, TypeSafe raised $870 million: investors are betting on the pioneer. The value is moving towards what cannot be copied as quickly: probabilities that can be trusted, measured by someone other than the vendor.

Second effect: AI splits in two, with fast models that decide at each step and large models called only to write or reason. Marc Brooker, the Amazon engineer behind Strands Decider, sees in them “a perfect decider for a workflow step” (TechCrunch). Each turn of an agent’s loop can go through a quick check, a precaution that the case of OpenAI’s agents that hacked Hugging Face makes very concrete. And the router that chooses the model becomes a model itself.

Third effect: fine-tuning becomes a product. Jev serves the same weights to all its customers and does not train on their data: you adapt it through the request, according to its documentation. Cloudflare takes the opposite approach with a fine-tuning service for Clef: your traffic, captured by its AI gateway, becomes a dataset, then the model is retrained and redeployed on its platform, at first with its engineers, later as self-service. It warns itself that a fine-tuned model can lose in versatility what it gains in accuracy on its domain.

The fourth effect, the most important in our view, is in the name. Jev owes it to William Stanley Jevons, the economist who explained in 1865, in The Coal Question, that savings in coal do not reduce its consumption:

TypeSafe embraces the parallel: every order-of-magnitude drop in the cost of intelligence would open up orders of magnitude more use cases. Cheaper decisions will therefore not reduce the total AI bill; they will multiply automated decisions. This is the mechanism we described in connection with the energy and water consumed by AI.

What we think

The idea is sound, and we already apply it: separate what decides from what writes. Our internal routing table gives sorting, classification and intent routing to a very small model, with an open model on our own server as a fallback, and a written rule: the router itself must cost next to nothing. Decision models aim at exactly that slot. They will get in through our benchmark, not on their word. Their real novelty, in our view, is the probability: if it can be trusted, it lets you set a threshold on an informed basis.

But cheaper does not mean necessary. TypeSafe says so itself in its guide: keep deterministic work in code, which is “reliable and cheap”, and avoid agent loops when a software workflow can do the same job. A written rule pays for no model call and is never wrong about what it covers; a decision model does not replace it, it replaces a large model used for sorting. Not every automation needs AI, and putting it everywhere without need multiplies costs and makes projects drift. That is the principle behind our n8n automation, with or without AI offer: AI only where it pays off, and a decision model such as Jev when free text has to be sorted. Our guide to automation without AI helps you spot the automations that can do without it.

That leaves dependency. Jev is a young, proprietary product, hosted in the United States and trained first in English, whose price has not yet stood the test of time. Its $870 million gives it room to last, not a guarantee that the price will hold: TypeSafe admits it cannot prove that the price is not subsidised. Clef reduces that risk: it uses the same API, and its open weights can be installed on your own servers. Keep the decision behind your own interface, so that changing supplier remains a configuration change.

What would change our mind: independent evaluations, not published by the vendors, showing that these probabilities remain reliable on business data in French, and prices that hold for a year.

How do you implement a decision model?

Here is the approach we follow. It starts with the work, not with the model.

1. 01Audit the decisions that keep coming back.

   We map the processes with the people who do the work, take stock of the tools and what they cost, check which ones can be connected (an API, an export, nothing at all), and gauge the team’s reservations and appetite, as well as its training needs. The repetitive decisions come out of it, with their volume and the cost of an error. That is the purpose of our digital and AI maturity audit, which is used to rank your projects by impact and feasibility, with a written GO/NO-GO, in 2 to 4 weeks for €3,500. It is an investment, and it makes the most sense when several dozen people repeat the same decisions.

2. 02Pick the right step on the staircase, decision by decision.

   A rule if the decision can be written down (an amount, a date, a status). A decision model for the fuzzy cases: a freely written email, an agent action to judge. A large model when something has to be written or reasoned through. A human below the threshold, and always for a decision that affects a person.

3. 03Design the questions

   Following the principles in the next section.

4. 04Test, then observe.

   The benchmark compares several models on 20 to 30 real samples, a decision model alongside conventional models: enough to choose, not enough to guarantee an error rate. The chosen model then works in observation mode: it decides alongside the team, without acting, until each confidence band has enough cases. To measure an error rate of about 10% to within 3 points, at the usual 95% confidence level, you need nearly 400 cases in the same band (the standard calculation of a confidence interval on a proportion). Our guide to evaluating an AI model before production sets out the method.

5. 05Plug the model in behind your own interface.

   The business code only talks to that interface, which holds the budget and the ceiling for each use, the fallback in case of an outage and the cost log: changing model does not touch the business code. The thresholds are set from the observation phase, and are higher for an irreversible action.

6. 06Train the team.

   The people who take over the cases below the threshold learn to read a probability and to correct the model. Each correction is recorded: it is the material that will be used to recalibrate.

How do you design the questions for a decision model?

A decision model is only as good as its questions. TypeSafe’s guide boils down to a few principles that we adopt as our own:

• Atomic questions. One question, one property. “Is this message an urgent complaint about an order?” hides three judgements; three short questions bring them out, and the code combines them. This is, according to TypeSafe, probably the most important principle in its guide.
• Explicit criteria. For a choice, each option says what it covers and what belongs to another, with a few examples. This is where your business enters the model.
• Only the useful context. A state cluttered with unnecessary information loses accuracy, and TypeSafe advises against counting on what the model already knows when the up-to-date information is in your database.
• Calculation left to the code. Jev counts poorly and reads dates as text, according to its page of known weaknesses. An amount, a deadline or a number of items is calculated before the model is called.
• Option order tested. Jev sometimes leans towards the first option in a list: check that shuffling the options does not change the answer. Cloudflare says it trained Clef on data where the order of the fields and the form of the questions vary; the test is still useful.
• Thresholds matched to the risk. TypeSafe suggests three zones: act when confidence is high, ask for confirmation or review in between, hand over when it is low. In its banking example, displaying a balance only needs to clear a floor of 50%, whereas approving a transfer requires more than 90%, and even then with a confirmation. Start cautious and adjust on your data, advises its page on confidence.

An example. An online shop receives: “My order still hasn’t arrived, and I have been charged twice.” A single question, “Which department does this message fall under?”, would force a choice between delivery and billing. It is better to ask, in the same request:

1. “Does the message report a delivery delay?” (true or false)
2. “Does the message report a payment problem?” (true or false)
3. “What is the customer’s level of irritation?” (a score on four levels)

The three questions are about the same state and are evaluated in parallel: breaking the question down adds no round trip, TypeSafe points out. Before the call, the code has found the order and counted the debits in the database: the model does not have to guess. Then the code opens one task for delivery and one for billing, and gives the case priority if the irritation is high and the confidence sufficient.

Where should you run a decision model?

Three options, to be chosen according to your data, your volume and the response time the use can tolerate:

The vendor’s API is the simplest way to start: nothing to install, and its limit of 80 requests per second is far above the million monthly decisions in the cost example below, which works out at fewer than one a second on average. It suits public data. For internal data, our sensitivity grid calls for a provider that retains nothing: at TypeSafe, zero retention is presented as an option for enterprise customers, and the data processing agreement provides for the European Commission’s standard contractual clauses for transfers. Personal and regulated data stay self-hosted. For a European company, US hosting is open to question, as we explain in connection with dependence on the American cloud; OpenAI, incidentally, offers to process the requests of its Decisions API in Europe, subject to conditions and for an extra charge, according to its documentation.

Workers AI runs Clef on the GPUs that Cloudflare spreads across its network, as close to the user as possible, alongside large models hosted in the same place. The Clef page states a context of 65,536 tokens and the ability to read text, JSON, images and video; the Clef-omni page adds sound, and Clef-flash is limited to 24,576 tokens since its price cut. The free allowance on Workers AI, 10,000 “neurons” a day, covers nearly 460,000 input tokens a day with Clef and nearly 2.9 million with Clef-flash, which is about 300 decisions of 1,500 tokens in the first case and nearly 2,000 in the second; beyond that, the paid Workers plan charges $0.011 per 1,000 neurons. If your data must stay in Europe, check where inference runs: the model pages do not say.

Your own server keeps everything in-house. Strands Decider 2B even runs on a CPU alone, according to its model card, which puts it within reach of an ordinary server, with a speed to be measured. Clef-flash needs a GPU with at least 41 GB of memory, and Clef 85 GB, a Cloudflare product manager told The Register, for one request at a time with a full context. At Scaleway, a 48 GB L40S GPU rents for about €1,073 per month excluding VAT in Paris, according to its price list consulted on 11 October 2026. For comparison, $1,000 buys about 26 billion input tokens read by Clef-flash on Workers AI, more than 17 million decisions of 1,500 tokens: below such volumes, a server of your own is justified by the sensitivity of the data more than by the price. A GPU rented in Europe corresponds to level 3 of our four levels of private AI, a machine on your premises to level 4. We already run open models with Ollama on our own server, in production, for sensitive data.

How much does a decision model really cost?

Per call, very little. Let us take the assumptions from our guide to the cost of AI in business: 1,500 input tokens per decision and a 50-token answer, at the prices consulted on 11 October 2026, for an equal number of tokens (each model splits text in its own way).

Claude Haiku 5.5, Anthropic’s entry-level model, which the vendor intends for classification, extraction and routing, is counted at $0.10 per million input tokens and $0.50 per million output tokens, on Anthropic’s price list. OpenAI’s price is its base rate, excluding the surcharge for processing in Europe. On Workers AI, the free allowance covers the first column, which represents fewer than a hundred decisions a day. And asking three questions about the same message does not triple the bill: Jev reads the state once and evaluates all the questions in parallel, according to its documentation.

For an SME that sorts 2,000 emails a month, the gap is counted in cents, whereas a misfiled email costs the time of the person who catches it. Against the most recent small language models, price is in any case no longer enough to tell them apart: Claude Haiku 5.5 costs only about three times as much as Jev, and less than Clef. The gap only really matters at very high volume, that of agents that have every one of their actions checked: at a hundred million decisions a month, Jev would cost about $6,300 and Haiku 5.5 $17,500. The rest comes down to speed, and to a calibrated probability that language models do not provide out of the box.

The combination weighs more than the unit price. Take a textbook example, where each step of the staircase receives only what the previous one did not settle:

Total: about $32 a month, against $175 if everything went through Haiku. The proportions are invented for the example; only yours count, and the audit is there to measure them.

The real cost is elsewhere:

• Set-up. The audit, the benchmark and the design of the questions weigh far more than the model calls at low volume.
• Human time below the threshold. A threshold that is too cautious sends too many cases to the team; one that is too bold lets errors through. The right setting can be read from the decision log.
• Monitoring. A workflow that fails silently costs more than an expensive model: every failure must trigger an alert, and every decision must leave a trace. Our guide to monitoring your n8n workflows shows how to be warned in time.
• The Jevons effect. Decisions that cost almost nothing multiply, and with them the errors to watch. That is why every automation we deliver ships with its cost counter.

How do you monitor a decision model once it is live?

A decision model is not something you install once and forget. Four habits keep it reliable:

1. A decision log. For each decision: the model version, the options, the probabilities, the threshold applied, what the code did, and the human correction if there was one. This log makes it possible to recount errors by confidence band, and it prepares you for what the AI Act will require of high-risk uses: activity logging and human oversight.
2. A pinned version. Jev’s documentation advises pinning the version your thresholds were tuned against, then moving to the next one at your own pace, after a new test.
3. A regular review. A growing share of cases below the threshold, or more human corrections, signals that your data has changed: new products, new requests, new vocabulary. Our “AI at the right cost” pack includes a monthly report and a quarterly review of the routing table, and decisions go through it like everything else.
4. Specialisation, when the volume justifies it. Jev cannot be fine-tuned: you adapt it through the content of the state and the criteria of the questions. Clef can be, with Cloudflare’s service. Strands Decider publishes its training recipe, about 11 hours on an RTX 3090 card, and its model card advises training the model on a rubric close to yours. The corrections recorded by your team then become training data, with the trade-off Cloudflare points out: a fine-tuned model can lose versatility.

Where to start

Start from the decision that comes up most often in your automations, such as sorting customer support tickets or checking an action. Measure what it costs and how often it gets it wrong, then ask yourself whether a rule is enough; if not, a decision model deserves a test, with a threshold and a human below it. Our Business AI: private and at the right cost section follows these choices, and our “AI at the right cost” pack is there to compare models on your own content, with cost measured and quality judged, before you decide which one to keep.

Frequently asked questions

Sources

1. Jev’s documentation (new tab)docs.typesafe.ai
2. The Clef announcement (new tab)blog.cloudflare.com
3. Its design guide (new tab)docs.typesafe.ai
4. The Register (new tab)theregister.com
5. The Strands Decider model card (new tab)huggingface.co
6. SiliconANGLE (new tab)siliconangle.com
7. Jev (new tab)docs.typesafe.ai
8. Workers AI (new tab)developers.cloudflare.com
9. OpenAI (new tab)developers.openai.com
10. Perplexity (new tab)docs.perplexity.ai
11. Are published on Hugging Face (new tab)huggingface.co
12. Perplexity’s model (new tab)huggingface.co
13. TechCrunch (new tab)techcrunch.com
14. Its second announcement (new tab)blog.cloudflare.com
15. The Register (new tab)theregister.com
16. The 2022 paper on InstructGPT (new tab)arxiv.org
17. TechCrunch (new tab)techcrunch.com
18. TypeSafe’s use-case map (new tab)docs.typesafe.ai
19. TechCrunch (new tab)techcrunch.com
20. The case study published by TypeSafe (new tab)typesafe.ai
21. Jev Router (new tab)openrouter.ai
22. The European Commission (new tab)digital-strategy.ec.europa.eu
23. GDPR (new tab)eur-lex.europa.eu
24. Its announcement (new tab)typesafe.ai
25. OpenChamber (new tab)openchamber.dev
26. According to Vercel (new tab)vercel.com
27. The Coal Question (new tab)econlib.org
28. Its page of known weaknesses (new tab)docs.typesafe.ai
29. Its page on confidence (new tab)docs.typesafe.ai
30. The data processing agreement (new tab)typesafe.ai
31. The Clef page (new tab)developers.cloudflare.com
32. The Clef-omni page (new tab)developers.cloudflare.com
33. Its price list (new tab)scaleway.com
34. Anthropic’s price list (new tab)platform.claude.com