Decision models are popping up like mushrooms after the rain

What a decision model is

An agent spends most of its time making small calls. Is this ticket urgent? Which tool next? Is this shell command dangerous? Is the task done? Until last month the default was to send each of those to a general LLM and parse whatever came back.

A decision model takes the situation and a set of questions whose answers you define in advance (yes or no, one of these options, a score on this scale), and returns a probability for each answer. No text, no parsing, no fourth option the model made up. Your code keeps the branches and the model supplies the fuzzy judgment. Cloudflare's definition is "A decision model makes classifications to help agents decide how to act, based on certain probabilities" ^[2]. TypeSafe's Jev started the wave on 15 September 2026 ^[1].

The rain

The numbers are counted from our decision model listings, and each listing links the vendor pages it was graded from ^[20]. Here's the order they arrived in. Jev (15 September), Decider, Laya, Kev, Liquid d1, Nace's Drex, Vela 2.0, then on 1 October alone Cloudflare's Clef, Perplexity's decider and AWS's Strands Decider ^{[2] [11]}. OpenAI's Decisions API came on 6 October, Celeris-1 on 8 October, and on 9 October Microsoft-Decision-1 and Clef-omni, which takes audio and video as well ^{[5] [3]}. This week Nace published the Drex 1.5 weights, under a licence that needs a commercial agreement once a business passes $1 million in revenue or funding ^[16].

Why so many, so fast? Because they're cheap to make. Most of these are a small open model post-trained to score answers instead of writing them. Several start from Qwen3.5, Microsoft's included (it's post-trained from Qwen3.5-9B) ^{[5] [11]}.

You can make your own. Unsloth's guide puts the same kind of decision head Cloudflare uses for Clef on Qwen3.5-0.8B and trains it with LoRA. On its typed-decisions set, accuracy goes from 36% to 73%, on 4 GB of VRAM in 42 minutes. The quick version, 60 steps on Qwen3.5-4B, reaches 76% in ten minutes on a single L4 ^[7]. When a home GPU can make one over lunch, every lab with a GPU cluster will have one by Friday.

And they're cheap to run. Every hosted one charges only for input, between $0.02 and $0.24 per million tokens, and nothing for output.

Prices are from each vendor's pricing page on 10 October 2026 ^{[10] [4] [19] [18] [6] [14] [17] [8]} and change often. The decision models shortlist has the current ones, and the head-to-heads put any two side by side. Microsoft-Decision-1 against Jev is the closest pair on price, and Clef against Laya was the most-read page on the site in our first week.

Everyone wants to see your decisions

This is the part I care about. A prompt to a chatbot is noisy. A decision request is the opposite. It names the situation, the options the business is choosing between and what it will do with the answer. Route this customer to billing or engineering. Approve this refund or not. Pick supplier A or B. Is this transaction fraud? Collect enough of those and you know how a company runs, what it's choosing between and where it's unsure.

That's why the big names all want a model in this slot, at prices that can't make much money on their own.

• Microsoft says it will rebase Decision-1 on its own MAI models and on OpenAI's, and it sells it through Foundry, where your decisions sit next to the rest of your Azure bill ^[5].
• OpenAI runs its Decisions API on gpt-6-luna. Its docs say API data isn't used for training unless you opt in, abuse-monitoring logs are kept for up to 30 days, and Zero Data Retention is available to eligible customers ^{[8] [9]}.
• Cloudflare gives the weights away under Apache-2.0 and says "we don't read, store, or train on your requests or responses" (fine-tuning aside). The same launch post points out that AI Gateway can "automatically create a dataset of requests for your use case" and sells fine-tuning on top. Your decisions, turned into your dataset, turned into a paid service ^[2].
• Perplexity sells search to agents. In my view a decision model is the step right before an agent picks what to buy or read, which is a good place for a search company to stand ^[10].
• AWS ships Strands Decider as open source, built for agents made with its Strands SDK. Free, and it keeps your agent's habits close to AWS ^[11].
• Liquid AI's terms (30 September 2026) grant it "a worldwide, non-exclusive, royalty-free license to use Your Content to develop and improve Liquid AI's models", products and services, training included. Its privacy policy says it sells or shares personal data subject to an opt-out ^{[12] [13]}.
• TypeSafe says Jev isn't trained on customer requests. Its customer agreement also says it may derive telemetry from customer data "in perpetuity", and that it "may Process Telemetry without restriction" ^{[14] [15]}.

None of them says it will sell what it learns from your decisions, and I'm not saying they do. But look at the incentives. A model that costs four cents per million tokens is a way in, not a business. The business is the position. Whoever answers "which one?" for millions of agents knows what is being chosen and can be paid by whoever wants to be chosen. That can be routing fees, ads that look like answers, data products about what companies are deciding, or simply a better model next year trained on everyone's edge cases. In my opinion that's where the money will be.

We're not outside this. letme routes agents to tools and we plan to earn a referral fee from vendors, which is the same position. That's why we publish that ranking can never be bought, and why letme will say which vendors have a deal with us before there's money in it.

What I'd do

• Run it yourself if the decision is sensitive. Ten of the fourteen publish weights. Laya, Decider, Kev, Vela, Strands Decider, GLiClass, Clef and pplx-decider's are Apache-2.0; Liquid's and Nace's come with commercial limits. The shortlist shows which run locally.
• Read the data-use line before the price. Every decision model listing has what the vendor says about training and retention, with the source. Check it before you send it your refund queue.
• Ask for zero retention where it's offered. OpenAI offers it to eligible customers ^[8].
• Keep the questions generic. Send the facts the decision needs, not the customer record around them.

We'll keep listing them as they land. At this rate there'll be another one before you finish reading this.

References

Read on 10 October 2026. Quotes are as published, and benchmark and speed figures are the vendors' own.

1. [1]TypeSafe, Introducing System One models and Jev, 15 September 2026. Source for Jev's launch. typesafe.ai
2. [2]Cloudflare, Clef: open-weight decision models, 1 October 2026. Source for the definition of a decision model, the promise not to read, store or train on requests, and AI Gateway's datasets of requests. blog.cloudflare.com
3. [3]Cloudflare, Clef gets faster, cheaper and multimodal, 9 October 2026. Source for Clef-omni and the Clef-flash price cut. blog.cloudflare.com
4. [4]Cloudflare, Workers AI pricing. Source for the Clef and Clef-flash prices. developers.cloudflare.com
5. [5]Microsoft, Microsoft-Decision-1 in Microsoft Foundry, 9 October 2026. Source for the launch, the Qwen3.5-9B base and the plan to rebase on MAI and OpenAI models. commandline.microsoft.com
6. [6]OpenRouter, Microsoft-Decision-1. Source for its price of $0.042 per million input tokens and nothing for output. openrouter.ai
7. [7]Unsloth, Train your own decision model. Source for Qwen3.5-0.8B going from 36% to 73% on typed-decisions on 4 GB of VRAM in 42 minutes, and Qwen3.5-4B reaching 76% in 60 steps and ten minutes on an L4. unsloth.ai
8. [8]OpenAI, Decisions API guide. Source for gpt-6-luna, the $0.10 price and Zero Data Retention for eligible customers. developers.openai.com
9. [9]OpenAI, Data controls in the OpenAI platform. Source for no training on API data without opting in, and abuse-monitoring logs kept for up to 30 days. developers.openai.com
10. [10]Perplexity, API pricing. Source for pplx-decider's price. docs.perplexity.ai
11. [11]Strands Agents, Introducing Strands Decider, 1 October 2026. Source for its release, its Qwen3.5-2B base and its place in Strands. strandsagents.com
12. [12]Liquid AI, Terms and conditions, 30 September 2026. Source for the licence to use customer content to develop and improve its models. liquid.ai
13. [13]Liquid AI, Privacy policy. Source for selling or sharing personal data subject to an opt-out. liquid.ai
14. [14]TypeSafe, Models. Source for Jev not being trained on customer requests, and its price. docs.typesafe.ai
15. [15]TypeSafe, Master Customer Agreement. Source for telemetry derived from customer data in perpetuity and processed without restriction. typesafe.ai
16. [16]Nace, Drex 1.5 on Hugging Face and its licence. Source for the weights and the commercial licence above $1 million in revenue or funding. huggingface.co
17. [17]Nace, Drex pricing. Source for the $0.05 price. console.nace.ai
18. [18]Celeris, Pricing. Source for Celeris-1's price. docs.celeris.ai
19. [19]Liquid AI, Decision models. Source for Liquid d1's price and the d1-3B weights. docs.liquid.ai
20. [20]Anchor Terminal, decision model listings, shortlist and head-to-heads. Source for the counts, launch dates and licences, each listing with the pages it was graded from. anchorterminal.com