Jev AI: The New AI That Doesn't Chat, It Decides ⚡

Most AI models love to talk. Ask ChatGPT or Claude a question and you'll get a thoughtful paragraph (or five).

Jev AI is different. It doesn't write, chat, or explain. It makes decisions, fast. 🎯

Launched on September 15, 2026 by San Francisco lab TypeSafe AI, Jev is built for a simple idea: a lot of AI work isn't a conversation. It's a quick judgment call inside software. Let's break down what Jev is, why people are excited, and where it still has room to grow.

🤔 So, What Exactly Is Jev AI?

Think about questions like these:

Is this customer email urgent?
Should this ticket go to billing, tech, or sales?
Is this message a jailbreak attempt?

Your software doesn't need an essay for any of them. It needs a clean answer it can act on right away.

That's Jev's whole job. You give it some information plus a few questions, and it hands back answers like this:

{ "choice": "technical", "probabilities": { "billing": 0.08, "technical": 0.85, "sales": 0.07 }, "confidence": 0.82 }

No paragraphs. No parsing. Just a decision your code can use. ✅

💡 The easy way to picture it: If ChatGPT is a coworker you chat with, Jev is a super-smart switch hidden inside your app. Founder Diogo Almeida describes it as "more like a database than a coworker."

👩‍🔬 The Team Behind It

TypeSafe AI isn't a random startup. It comes with some serious credentials:


🧠 Founder & CEO	Diogo Almeida, ex-OpenAI researcher and co-inventor of RLHF, the technique behind ChatGPT
🤝 Co-founders	Erik Gafni and Sasha Sheng
💰 Funding	$40 million seed round led by DCVC
📍 Based in	San Francisco (founded 2024)

There's a nice twist here. Almeida helped teach AI how to talk to people. Now he's betting the next big leap is AI that works quietly for software. In his words, "people can't be the only consumers of intelligence."

🏷️ Fun Fact: Where the Names Come From

"System One" comes from Daniel Kahneman's book Thinking, Fast and Slow:

🐇 System 1: fast, instinctive gut-checks
🐢 System 2: slow, careful reasoning

Most modern AI models are racing toward System 2. Jev proudly picks System 1.

"Jev" honors economist William Stanley Jevons. He noticed that when steam engines got more efficient, people ended up burning more coal, not less. TypeSafe thinks AI will follow the same pattern: make intelligence cheap enough, and people will find a thousand new uses for it. 🔥

⚙️ How Does It Work?

The simple version 🧩

A regular AI model is like someone answering a quiz by writing a full letter for every question.

Jev is like an expert who fills in the whole multiple-choice sheet at once and notes how confident they are in each answer.

The slightly nerdy version 🤓

Three things make Jev tick:

🎓 A new training method, RLCD. Short for Reinforcement Learning for Calibrated Decisions. The goal is honest confidence: when Jev says it's 90% sure, it should be right about 90% of the time.
⚡ Parallel answers. Chatbots generate text one token at a time. Jev answers all your questions in a single pass, which is why adding more questions barely slows it down.
🧱 Three simple building blocks:

Building block	What it does	Example question	What you get back
🎯 Choice	Picks one option from a list	"Which team handles this?"	`"returns"` + a confidence score
📊 Score	Rates something on a scale you define	"How upset is this customer?"	`1.6` on your scale + a confidence score
✅ Noul	Gives a yes/no probability	"Is this a refund request?"	`0.98` (98% likely yes)

Because every answer must match a format you define in advance, Jev can never return something your code doesn't expect. That's the "type-safe" part of TypeSafe. 🛡️

🥊 Jev vs Traditional LLMs

	💬 Traditional LLMs	⚡ Jev AI
Output	Free-form text	Typed, structured answers
Speed	Seconds to minutes	~70–500 milliseconds
Confidence score	Often unreliable	Built into every answer
Format errors	Possible	Impossible by design
Explains itself	✅ Yes	❌ No
Best at	Chat, writing, coding	Sorting, scoring, routing, safety checks

🤝 Not a rivalry. Jev isn't here to replace ChatGPT or Claude. The smartest setups will likely use both: an LLM for the creative, open-ended work, and Jev for the thousands of tiny decisions around it.

💸 The Headline Numbers

This is where jaws dropped. 😮

Price: $0.042 per million input tokens (that's $42 per billion)
Output: Free. Yes, really.
Speed: 70–500 ms per response, which TypeSafe says is 40–200× faster than frontier LLMs on these tasks
Demo showdown: Jev answered in 0.114 seconds, while GPT-5.6 Terra took 8.566 seconds
Cost comparison: reported to be about 238× cheaper than Claude Fable 5.1

📝 Fair note: these figures come from TypeSafe and launch coverage. Independent benchmarks are still on the way.

🚀 What Can You Build With It?

🎫 Smarter support Route tickets, spot angry customers, and flag urgent issues instantly.

📊 Big data, small bill Run the same questions across millions of documents to turn messy text into clean, usable data.

🛡️ AI guardrails Screen chatbot inputs and outputs for jailbreaks, harmful content, or made-up citations, all in milliseconds.

🤖 Human + AI teamwork Let Jev act automatically when it's confident, and hand the tricky cases to a person when it isn't.

🎮 Yes... it plays Doom TypeSafe's favorite demo has Jev playing Doom in real time from a text description of the game. It makes about 10 decisions per second for roughly $7 an hour. Another demo has it "Wikiracing" between Wikipedia pages by picking the right link from hundreds of options.

👨‍💻 A Quick Peek at the Code

Here's how a developer might triage a customer email with Jev's Python SDK:

from typesafe_sdk import Choice, Noul, Score, TypeSafeClient

client = TypeSafeClient()

result = client.system_one(
    state="My order arrived broken. Second time! I want a refund today.",
    questions={
        "team": Choice(
            instructions="Which team should handle this",
            criteria={
                "returns": "Damaged or unwanted items",
                "shipping": "Delivery problems",
                "account": "Login or payment issues",
            },
        ),
        "mood": Score(
            instructions="How upset the customer sounds",
            criteria=["Calm", "Unhappy", "Very angry"],
        ),
        "wants_refund": Noul(instructions="The customer wants a refund"),
    },
)

team = result.answers["team"]
queue = team.choice if team.confidence >= 0.6 else "human_review"
print(f"Send to: {queue}")

The golden rule: Jev makes the fuzzy calls, and your code makes the final decision. 🧠➡️💻

(Jev is in early access, so check the official docs for the latest syntax.)

⚖️ The Honest Part: Limitations

Every shiny new tool deserves a reality check. Here's what critics are pointing out:

❌ "Can't hallucinate" ≠ "can't be wrong." Jev can't invent answers outside your options, but it can still pick the wrong one. The Register called the comparison a bit unfair for exactly this reason.
🧪 The speed trick may not be unique. Engineer Sean Goedecke argues you can get some of the same speed from regular LLMs with clever tricks. He still expects a purpose-built model to do better.
🐇 No deep thinking. Jev can't reason step by step, so big questions need to be broken into smaller ones.
🤐 No explanations. You get numbers, not reasons.
📋 Company-made benchmarks. TypeSafe built its own evaluations and openly admits they may be biased.
⏳ Early access only. For now, you'll need to join the waitlist.

🎯 The Bottom Line

Jev AI is a refreshing reminder that AI doesn't always have to talk.

By trading chatty text for fast, cheap, reliable decisions, TypeSafe is betting on a future where intelligence runs quietly inside the apps we use every day, much like a database does today.

👔 For businesses: a new way to automate the small judgment calls that eat up time.
👩‍💻 For developers: a fresh building block worth testing.
🙋 For everyone else: a peek at a future where the most useful AI might be the one you never chat with. ✨

👉 Want to try it? Join the waitlist at typesafe.ai.

❓ Quick FAQ

What is Jev AI? A "System One" AI model from TypeSafe AI that returns fast, structured decisions with confidence scores instead of generating text.

Who made Jev AI? TypeSafe AI, founded by former OpenAI researcher Diogo Almeida along with Erik Gafni and Sasha Sheng.

How much does Jev AI cost? $0.042 per million input tokens. Output is free.

How fast is Jev AI? About 70–500 milliseconds per response, according to TypeSafe.

Is Jev AI a ChatGPT alternative? Not really. It's a teammate for chatbots, built for quick decisions inside software rather than conversation.

Can I use Jev AI today? It's in early access. You can join the waitlist on typesafe.ai.

📚 Sources

Figures are as reported at launch (September 2026) and may change as Jev moves beyond early access.