System one models decide between fixed options instead of generating text, and Jev is the first public one. It runs in milliseconds, costs a fraction of a chat model per token, and hands the chosen option back to the program that asked.
What System One Models Actually Do
System one models are decision models: they select one option from a fixed list instead of generating free-form text, and Jev from TypeSafe AI is the first public release built on that approach. That constraint is the whole design. Because the answer must come from a list the caller supplied, the model never composes a sentence, and the program that asked receives a value it can act on immediately.
The distinction matters more than it first appears. ChatGPT and Claude general assistants used for writing, research and multi-step reasoning. A system one model is used at the point where the choices are already known and something has to pick fast: which model handles a subtask, whether a game character jumps, which element a browser agent clicks next.
Jev returns the chosen option together with a confidence score. The caller decides what to do with that score, so the model itself does not have to be trusted blindly; software can route low-confidence picks to a slower path.
Jev vs LLMs: A Same-Dimensions Comparison
Jev and a general-purpose LLM differ on output shape, latency and cost before they differ on raw intelligence, so comparing them as competitors misses the point. The table below uses only the dimensions the launch material and hands-on demonstration actually address. Feature-level detail comes from TypeSafe AI's console documentation, and the task-level behavior comes from a build walkthrough published on 22 September 2026.
| Dimension | Jev (system one model) | ChatGPT / Claude (LLMs) |
|---|---|---|
| Output | One option from a fixed list, plus confidence | Free-form text, generated token by token |
| Typical response time | About 70-500 ms per decision | Seconds per response, longer for reasoning |
| Cost basis | Per-token pricing reported by the vendor | Per-token pricing, higher per call |
| Best fit | Repeated small decisions inside a program | Writing, planning, research, complex problems |
| Failure mode | Wrong pick inside a known option set | Plausible text that needs checking |
| Integration | Called from code, including via Claude Code | Chat interface, API, document tools |
A bigger model can hold the main plan while Jev handles the smaller decisions underneath it. That division of labor is the practical answer to the comparison, and it explains why the two are usually deployed together rather than as substitutes.
How RLCD Makes Decision Making Faster
RLCD is the method TypeSafe AI describes behind Jev: the model is given a fixed set of possible answers and chooses among them rather than generating whatever it wants. Restricting the output space is what makes the result faster, more predictable and easier to control than sampling from a language model's full vocabulary.
The demonstration video shows this in practice. A request to write a blog post is split into three jobs, and Jev decides where each goes: the article to Claude, the cover image to an image model, the tagging to a cheaper model. Each routing choice carries its own confidence score.
In a reaction game where a character can stay still, move left, move right or jump, Jev re-decides as blocks approach. The published walkthrough reports it making around 10 decisions per second in one test, which is the range where a per-decision chat-model call would stall the game.
What Jev Costs and How Fast It Responds
TypeSafe AI's launch material quotes Jev at input tokens of $42 per billion tokens with output tokens free, and response times around 70 to 500 milliseconds. Both figures are vendor-reported at launch and have not been independently reproduced in the material reviewed here, so treat them as claims from the company selling the model.
The two numbers pull in opposite directions from a normal model comparison. Latency is quoted per decision, not per conversation turn, and the cost structure assumes many small calls rather than one large one. A system that makes ten decisions a second is not paying for ten chat completions.
Per-option output also removes the parsing step. When a program already knows the set of valid answers, it does not need to interpret a sentence to find out which one was chosen.
Four Worked Examples From the Jev Test Build
The published walkthrough tests Jev in four places where a program needs a quick choice, and each one isolates a different reason to use it. Setup starts at the TypeSafe AI console, where the quick start guide offers a Claude Code command that is copied into a terminal and run; the TypeSafe skill is then loaded inside Claude Code.
AI traffic controller
A single request to write a blog post is split into three smaller jobs, and Jev routes each one: the article to Claude, the cover image to an image model, and the tagging to a cheaper model. Every routing choice carries a confidence score, so an expensive model is not used for work that does not need it.
Reaction game
The character can stay still, move left, move right or jump. As blocks approach, Jev keeps deciding, and the walkthrough reports roughly 10 decisions per second in one test. The same run also shows the reasoning: while the nearest block is far away the threat stays low and the character holds position.
Browser task
Asked to find the price of the pro plan, Jev makes one decision per step rather than one large decision. It reads the current page, checks the possible actions, picks one, and the browser executes it. The task ends with the price, $49 per month, after a short sequence of picks.
Smart home engine
A single instruction such as going to sleep expands into several decisions: curtains close, lights off, front door locks, TV off, thermostat to 20 degrees Celsius, security system on. Switching to movie night changes which of those decisions fire, so one human sentence becomes a set of discrete choices.
Where Jev Fits in an AI Agent Stack
Jev is best understood as a decision layer beneath a planner, not as a replacement for the models that do the planning. A large model can hold the overall task while Jev handles the small choices that would otherwise add latency and cost at every step.
That framing also sets the limits. Jev returns a pick from a list the caller defined, so it cannot invent the list or explain a decision in prose. If the options are not enumerable, or if the answer has to be reasoned out and written down, a general model remains the right tool. Tasks with a small, stable action space and a tight latency budget are where the design pays off.
The vendor's own positioning supports this reading: the launch material describes composable intelligence optimized for decisions rather than a general chat replacement.
Limitations and What the Demo Does Not Prove
Everything concrete about Jev's speed, pricing and task behavior in the reviewed material comes from TypeSafe AI or from a hands-on build published by one YouTube channel on 22 September 2026. That is vendor evidence and speaker experience, not independent measurement.
The public material does not show how Jev handles ambiguous or poorly specified option sets, what happens when none of the supplied options is correct, or how the confidence score is calibrated across domains. A reaction game and a smart home simulator are narrow settings, and performance there does not establish accuracy on tasks with long horizons or costly mistakes.
Anyone adopting Jev for production routing should build their own evaluation on their own option sets before trusting the confidence score to gate expensive decisions.
FAQ
- Are system one models the same as small language models? No. A system one model is defined by its output, not its size: it selects one option from a fixed set instead of generating text. A small language model still generates tokens, so it can write an answer but cannot guarantee the answer is one of your valid options.
- Is Jev faster than ChatGPT? TypeSafe AI quotes Jev at roughly 70 to 500 milliseconds per decision, while a general chat model typically takes seconds per response. The comparison favors Jev for a single structured choice; it says nothing about multi-step reasoning or writing quality, which are different jobs.
- Does Jev replace ChatGPT or Claude in an agent? Usually not. The demonstrated pattern keeps a large model for the main plan and uses Jev for routing and small step-by-step choices. Jev returns a selection and a confidence score rather than prose, so a general model still handles anything that has to be written out.
- How do you connect Jev to an existing agent? According to the build walkthrough, you open the quick start guide in the TypeSafe AI console, copy the provided Claude Code command into a terminal, then load the TypeSafe skill inside Claude Code so Jev is available to the agent.
- What does Jev cost? The launch material quotes input tokens at $42 per billion tokens with output tokens free. That figure is vendor-reported and should be rechecked against current pricing before you budget for it, since pricing on fast-moving AI products changes often.
Turning a Video Walkthrough Into a Written Guide
The Jev demonstration works because it walks through four concrete projects in order, and that order is exactly what a reader needs when they are deciding whether to add a decision layer to their own stack. Written explanations of the same material lose that sequence if they are padded with commentary, which is why the useful part of this article stays close to what was actually built and measured.
If you already explain technical work on video, the same sequence exists in your recordings: the setup step, the first working example, the failure, the fix. Skalablog takes a YouTube URL, transcribes the video, and generates a structured article from it, so that walkthrough you already spent an afternoon recording can reach readers who will never watch it.
Start at Skala Blog, paste the video link, and edit the draft that comes back. For teams building decision layers and agent tooling, CrazyStack is a useful reference point, as is the work shared by Dev Doido do canal do youtube on applied AI projects.
Fork this article
Start a new branch from the same video, shaped your way. You keep the credit; the original keeps the attribution.
A fork in another language is filed as a translation of this article, so the two pages point at each other. You can unlink it later from the editor.
0/240
You are creating
- Format
- For
- Language
- Source
- Your angle
You will be asked to sign in before it is generated.
Buy credits