Should You Move Claude's Busywork to Jev? — The AI Model That Can't Chat
On 15 September, TypeSafe shipped Jev: a model that can't write a sentence, only pick answers from a list you define, at a price TypeSafe says is up to 445 times lower. The arithmetic on TypeSafe's own numbers, what developers built with it in the first two weeks, and where it belongs next to the Claude you already pay for.
The Promise
- For decisions with a fixed menu and a volume problem, the arithmetic is hard to argue with: 4.2 cents per million input tokens, output free, roughly 56,000 support-ticket decisions for $1.
- Jev can't go off the menu. It returns only answers you defined, with a probability on each, which makes it a natural gate in front of, around or behind a larger model.
- Six early patterns already keep Claude in the loop: routing tasks between models, trimming context, screening prompts, gating agent tool calls, sorting test failures and grading agent answers.
The Risk
- The 445× figure measures how cheaply Jev agrees with GPT-6 Astra and Claude Fable 5.1, not whether any of them was right. Against Claude Haiku, the model most teams actually use for sorting, it is closer to 44×.
- Jev can't abstain. If none of your options fits, it picks the least wrong one, so every schema needs an 'unclear' option.
- Every decision is a threshold somebody writes into code, and in intrusion detection those thresholds quietly became policy that nobody owned.
- API only, hosted in the US, weights and parameter count unpublished. The customer agreement allows use of your data for telemetry and fraud monitoring with no end date; zero data retention is an enterprise option.
A model that can’t chat
On 15 September, TypeSafe shipped Jev. Its founder, whose research at OpenAI became the foundation for ChatGPT, wrote on launch day: “Models have been superhuman at chat for years. So where is all the automation?”
Every chatbot you’ve used writes its reply one word at a time. Jev skips the writing. You hand it data, an email or a log line, and a set of questions with the possible answers defined in advance. Is this urgent, yes or no? Which team owns it, billing or technical? Jev answers every question in a single call and puts a probability on each answer.
TypeSafe calls this “can’t hallucinate”. What it means is narrower. Jev can’t go off the menu. It can still pick the wrong item on the menu.
The arithmetic
Jev charges 4.2 cents per million input tokens, and output is free, so the cost of a decision is almost entirely the size of what you send. TypeSafe’s own worked example, a short support ticket with three questions at 426 tokens in, comes to roughly 56,000 decisions for $1.
The 445× headline comes from TypeSafe’s workflow tests, where the correct answer is defined as the average of what GPT-6 Astra and Claude Fable 5.1 said. It measures how cheaply Jev agrees with the big models. TypeSafe says as much on its launch page, which is to its credit.
Price the same ticket on Claude Haiku, at $1 in and $5 out, and it’s roughly 44 to one. Against Claude Fable 5.1 it’s about 440 to one. The headline holds when the comparison is the most expensive model on the market. Against the model you’d really use for sorting, it’s closer to 44. Still a large number. TypeSafe also says it can’t prove the price isn’t subsidised, so budget as if it might rise.
Where it goes: next to Claude
Developers keep landing in the same place. Jev goes in front of the model, around it or behind it.
In front: routing work between Haiku and Opus inside Claude Code (one builder saw his bill drop about 70% on 12 prompts, and says he didn’t measure quality), and LiteLLM’s relevance check, which drops low-value context before it reaches Claude.
Around: screening prompts and outputs for jailbreak attempts, and LangChain middleware that checks each tool call an agent is about to make and blocks the risky ones before they run. A log tells you what happened. A gate decides whether it happens.
Behind: sorting test failures before a frontier model reads 500 lines of error output, and Langfuse’s use of Jev as a judge. One line in the Langfuse guide is worth keeping: Jev can’t abstain, so always give it an “unclear” option.
The threshold is the policy
Every one of those jobs runs on a threshold. Jev returns a probability, and somebody writes the line of code that says: above this number, act without asking.
In my experience, this is where security tooling either works or fails quietly. For 20 years we’ve tuned intrusion detection by picking a threshold, and the threshold became policy. Nobody wrote it down, and nobody owned it. TypeSafe says Jev’s probabilities are calibrated. Test that on your own data, then name the person who owns the number before you ship.
Read the paperwork too. The product runs under TypeSafe’s customer agreement, not the website terms. It’s hosted in the US, which for European teams is a conversation with your data protection officer before it becomes a pilot. Pin the version.
If you’re paying a chat model to pick answers from a list, label a hundred of your own examples and run them through Jev before your next renewal. The launch page can’t tell you how it does on your data. An afternoon can.