Fast, cheap AI that makes decisions: what it is and how to try one
For small decisions made over and over, how fast the answer comes back matters as much as how smart it is, and a decision model answers for a fraction of a cent.
What the reel showed
Two snakes on the same 20 x 20 board chase the same food in the same order for 30 seconds. Every move is a live decision from Jev, TypeSafe's decision model, over OpenRouter. The snake moves one cell each time an answer lands, so the speed you see on the board is the decision rate and nothing else.
The left board runs at Laya's published response time, 38.4ms per decision. The right board runs at Jev's measured round trip over OpenRouter, 376ms at the median. In the same 30 seconds the fast board made 782 moves and ate 47. The slow board made 78 moves and ate 4.
The whole run was 860 live decisions and cost $0.017.
| Entrant | Tool | Time | Cost | Verdict |
|---|---|---|---|---|
| Laya lane (Jev's decisions) | Laya's published latency | 38.4ms per move | $0.015 | 47 food, 782 moves, 1 death |
| Jev | OpenRouter, System One | 376ms per move | $0.0015 | 4 food, 78 moves, 0 deaths |
Both boards use the same decision model; the left one runs at Laya's published response time of 38.4ms, the right at Jev's measured 376ms over OpenRouter.
What this means for your business
Jev and Laya are decision models. They do not write text. You give them a situation and a short list of options, and they hand back one option with a probability. That is the exact shape of most small decisions inside a business.
Think of routing an inbound lead to the right salesperson, tagging a support ticket, deciding whether an email needs a human today, flagging an order for review, or approving the next step in an automation. Each is a choice between a few known answers, made hundreds or thousands of times a week.
The reel shows why speed matters there. Ten times the decisions in the same half minute turned 4 into 47. In a business that is the difference between a queue that clears as items arrive and one that backs up. And it is cheap: Jev charges only for what you send it, output is free, and 860 decisions cost under two cents.
The skill is in writing the options. Plain up, down, left and right told the model very little. Describing what each move does, and naming the one that reaches the food, is what made it play. Your business rules work the same way: 'urgent' means little, 'the customer says their checkout is down' means a lot.
What I’d use
Start with Jev on OpenRouter. It is live, it needs no separate signup, only an OpenRouter key, and you pay per use. It is what I use for these tests.
Laya is worth following. Convai Innovations publishes it as open weights under the Apache 2.0 licence, built for speed, to run on your own hardware. Their own model card shows the base model scoring 0.362 on their typed-decisions benchmark and 0.766 once fine-tuned for that task, so plan on training it with your own examples. That is a project for someone technical, not an afternoon test.
The rule I use: when the output is words a customer reads, use a chat model like Claude. When the output is a choice, use a decision model.
Try it yourself
The quickest way to see a decision model work on your kind of problem. Plan on half an hour. The test itself costs well under a cent.
Create an OpenRouter account
OpenRouter gives you one account and one bill for hundreds of models, Jev included. Sign up and add a few dollars of credit. You pay per use.
Watch Jev work in your browser
OpenRouter's Jev Lab has live demos. Sign in and run Support message triage: it sorts 95 support messages against five yes-or-no questions and races a chat model on the same job.
Create an API key
This is what lets a script use your account. Treat it like a password: never paste it into a chat or a shared document.
Install Claude Code
Claude Code is Anthropic's AI agent that writes and runs code for you, so you do not have to. It is included with Claude Pro and Max, which is how I use it every day: sign in with your Claude account and only the Jev calls go on your OpenRouter bill. No Claude plan? Run it on your OpenRouter key with OpenRouter's Claude Code guide (linked below), or use Codex CLI on a ChatGPT plan instead.
Make a folder, add your key, paste the prompt
Create an empty folder. Inside it, make a text file called .env with one line: OPENROUTER_API_KEY=YOUR_OPENROUTER_KEY, using your real key. Start Claude Code in that folder and paste this.
I want to test Jev, TypeSafe's decision model, on OpenRouter. My OpenRouter key is in the .env file in this folder as OPENROUTER_API_KEY. Never print it. 1. Read https://openrouter.ai/docs/guides/community/jev-tutorial 2. Write 20 short, realistic customer emails for a business like mine: [describe your business in one line]. Mix sales enquiries, complaints, invoices, spam and newsletters. 3. For each email, call the Decisions API (POST https://openrouter.ai/api/alpha/decisions, model typesafe/jev-1.13) with one "choice" question: "What should we do with this email?" and these options: - reply_now: "A customer or prospect is waiting on us, is unhappy, or money is at stake." - later: "Useful but not urgent: updates, questions that can wait a day, internal notes." - bin: "Spam, cold sales pitches, newsletters nobody asked for." 4. Print a table: email subject, Jev's choice, its confidence, and how long each call took. Then the total cost from the usage.cost field. 5. Tell me which ones you would have sorted differently, and why.
Try it on your own inbox
Swap the samples for 20 real emails with names and addresses removed. Where Jev gets one wrong, rewrite the option descriptions, not the email. That is most of the craft.
Links
- OpenRouterOne account, every model, pay per useopenrouter.ai
- Jev 1.13 on OpenRouterModel id typesafe/jev-1.13, pricing and speedopenrouter.ai
- Jev LabLive Jev demos in the browseropenrouter.ai
- Jev tutorialYour first decision call, step by stepopenrouter.ai
- TypeSafe docsThe company behind Jevdocs.typesafe.ai
- Claude CodeAnthropic's coding agentclaude.com
- Claude Code with Pro or MaxRun it on your Claude subscriptionsupport.claude.com
- Codex CLIOpenAI's coding agent, signs in with a ChatGPT plangithub.com
- Claude Code on OpenRouterRun Claude Code on your OpenRouter keyopenrouter.ai
- LayaConvai Innovations' open decision modellaya.convaiinnovations.com
- Laya on Hugging FaceOpen weights, Apache 2.0huggingface.co
- Laya live demoConvai's demo on Hugging Facehuggingface.co
Want a decision model sorting your inbound leads?
Trying it takes an afternoon. Getting it right inside a real business, with your tools, your rules and your edge cases, is the part I do.