- Book: AI That Answers
- The series: AI in TypeScript — 5 books, from your first LLM call to agents in production — all five here
- Pocket Guides: AI Agents · LLM Observability · RAG · Prompt Engineering — the whole series
- My project: Hermes IDE | GitHub — an IDE for developers who ship with Claude Code and other AI coding tools
- Me: xgabriel.com | GitHub
You ask ChatGPT a question and watch the answer type itself out,
word by word. Claude does the same.
Picture an AI that never writes back. You hand it a situation and a
short list of possible answers. It picks one and tells you how sure
it is. Its maker says the whole exchange takes half a second or
less.
TypeSafe AI launched that model, Jev, on 15
September.
The Register
reports the startup has raised $40 million. Ask Jev for a birthday
message or a cover letter and it has nothing to give you, because
TypeSafe built it only to choose.
Diogo Almeida, TypeSafe's co-founder and CEO, announced Jev
on X. His post opened like this:
"After co-inventing ChatGPT, I kept asking myself: why have
superhuman chat models not led to AGI?" AGI, short for artificial
general intelligence, means an AI as broadly capable as a person.
The Register describes Almeida as a former OpenAI researcher and a
co-inventor of ChatGPT.
In the same post, he said he had "spent the last 2 years in
stealth", building without telling the public, and called the
result "a new type of frontier AI model", using the industry's label
for its most advanced systems.
Fast thinking, slow thinking
TypeSafe's homepage calls Jev the "first
public System One Model, optimized for automation". The
launch post
explains the name. It borrows from the psychologist Daniel
Kahneman, who described two modes of human thinking.
System 1 is the fast gut call. A ball flies at your face and you
duck before you have thought about it. System 2 is slow and
deliberate. It's the mode you use to compare two phone contracts
line by line.
Chat models like ChatGPT and Claude sit on the System 2 side of that
split. They reason in words and write out their answers. TypeSafe's
launch post, citing an outside benchmark site, puts the time
frontier models take from question to finished answer at 3 to 329
seconds. The slow end is more than five minutes. Jev sits on the
System 1 side. It makes one quick judgment and stops.
According to TypeSafe's docs
and a guide for programmers
who connect it to their apps, Jev can give only three kinds of
answer:
- Pick one. It chooses a single option from a list you supply, up to 255 options long.
- Rate it. It places something on a scale of 2 to 10 steps you define. For a support message, you might use "Low", "Medium" and "Urgent".
- Yes or no. It gives the probability that the answer is yes.
Pick-one and rate-it answers also come with a confidence number.
TypeSafe describes its output as "calibrated probabilities".
Calibrated means the numbers match reality over time, the way a good
weather forecast does. If a forecaster says 80% chance of rain on
ten different days, it should rain on about eight of them. When a
calibrated AI says it is 80% sure, it should be right about eight
times in ten.
The Register printed an example of Jev deciding which team should
handle a request: technical at 0.85, billing at 0.08 and sales at
0.07. That puts Jev's own estimate at about 85% that the request
belongs with the technical team. No paragraph comes back, so a
program can act on the answer straight away.
Nathan Flurry, co-founder of the software company Rivet,
wrote on X that "jev is just a
really smart switch statement". A switch statement is a piece of
code that works like a railroad switch, sending each item down one
of several tracks. He added that it's "like if 2016 ml classifiers
got 2026 levels of intelligence". Classifiers are older
machine-learning tools that sort things into labels, the way a spam
filter does.
TypeSafe's bet: a lot of AI never talks to anyone
TypeSafe is betting that a large share of AI work inside software is
a stream of small decisions that no person ever reads.
Vercel, a company whose cloud platform hosts websites and apps,
lists the kind of jobs Jev is meant for in its
product-update notes.
One is choosing an AI assistant's next move: keep going, try again,
ask the user or stop. Others include rating how urgent or risky an
action is before it runs and checking that an AI's output follows
safety rules.
If a company uses a chat model for those calls, each one means
waiting for a written reply, then having software dig the decision
out of the text.
Faster and cheaper, by TypeSafe's count
TypeSafe runs its own evaluations on
four business-style tasks: security incidents, customer service,
invoice processing and reviewing the work of AI agents, assistants
that carry out multi-step jobs on their own. Each test case is one
item run through a short set of questions. These are TypeSafe's own
numbers, not an independent test.
The comparison includes GPT-5.6 models from OpenAI, the company that
makes ChatGPT, and Claude models from Anthropic, the company that
makes Claude. The prices are per-use rates that businesses pay when
their software calls those models. A ChatGPT or Claude subscription
is priced differently.
On average, Jev took about 0.4 seconds per case. OpenAI's GPT-5.6
Luna took 12.9 seconds, and Anthropic's Claude Opus 5 took 37.8
seconds. A check that finishes in under half a second can review
every command an AI assistant wants to run, which is the job Vercel
tested it on. The
launch post
also shows Jev playing the video game Doom, choosing moves ten times
a second, for about $7 an hour by the company's count.
Cost per case was $0.0004 for Jev, $0.0033 for Luna and $0.1761 for
Opus 5. Scale that to a million cases and you get about $400, $3,300
and $176,100.
Jev scored 67.8% in the same tests. Luna scored 66.8%, and GPT-5.6
Terra, another OpenAI model, scored 67.9%. The top score was 74.1%,
from GPT-5.6 Sol, with Claude Opus 5 close behind at 73.1%. No model
came near 100%, so the useful comparison is between the models. Jev
lands in the middle of the pack, about six points behind the leader,
at a small fraction of the cost.
The widest gap was on invoices. Jev scored 61.8% on invoice
processing, against 79.1% for GPT-5.6 Sol.
Vercel tried it on a real job
Vercel also makes fx, an AI coding assistant that can run commands
on a developer's computer. A command is a typed instruction to the
computer, and the wrong one can delete files. Vercel's CEO,
Guillermo Rauch, wrote on X: "Default mode
in fx is auto, with a safety reviewer analyzing every command. That
reviewer runs on GPT Luna today."
He went on: "Jev is up to 18x faster (p95) and more accurate.
It's coming to @vercel AI Gateway and likely new default." The p95
figure looks at the slow end, the time that 95 out of 100 requests
come in under. Pranit, from Vercel's team,
said on X that Vercel ran the benchmark. As
of 17 September, Vercel hasn't published the data behind "more
accurate".
Rauch's word was "likely", and as of his post the fx reviewer still
ran on GPT Luna. A day after launch, Vercel's changelog said Jev "is
now available on AI Gateway", Vercel's own service for connecting
apps to AI models. That note mentions no waitlist. Getting Jev
straight from TypeSafe still means waiting for early access.
What Jev does not mean
Jev is no substitute for the chatbot you already use. Flurry said as
much: "jev does not replace gpt / claude". If you use a chatbot to
write, this launch changes nothing about your day.
"Zero hallucinations" is narrower than it sounds
People say an AI hallucinates when it states something false with
full confidence, like a made-up quote or a source that doesn't
exist. TypeSafe's homepage promises zero hallucinations for Jev. Its
launch post says that zero did not come from testing. The guarantee
covers the form of the answer.
Think of a multiple-choice exam. Jev will always circle one of the
letters printed on the page, and it will never write in an answer of
its own. It can still circle the wrong letter. Anthony Maio's
review of the launch on Substack
says the guarantee does nothing to stop the model "picking the wrong
option".
The numbers come from TypeSafe
Apart from the Vercel posts, every speed, cost and accuracy figure
here is TypeSafe's. The company's
launch post
says its evaluations may carry "some bias" because its own team
built them.
To grade answers, you need an answer key. TypeSafe built its key by
averaging the answers of two other AI models, GPT-6 Astra and Claude
Fable 5.1. A score of 67.8% means Jev agreed with that AI-written
key 67.8% of the time. Because other AI models wrote the key, a
model can disagree with it and still be right.
The homepage banner reads "193.6x Faster, 444.6x Cheaper", with a
footnote: "based on workflows for System One tasks". The launch
post calls those figures "the higher end of real world gains", and
says its side-by-side demo on the homepage is "highly simplified".
Some of the improvement in TypeSafe's tests doesn't need Jev at all.
Every chat model scored higher when the job was split into small
questions instead of one big request, and got cheaper and faster
too. OpenAI's Luna went from 51.9% to 66.8% that way, one point
behind Jev.
Does any of this get closer to AGI?
Almeida opened his announcement by asking why superhuman chat models
haven't led to AGI. Nothing TypeSafe has published tests that
question. Its evaluations cover business chores like customer
messages and invoices. Whether fast decision models are a step
toward AGI is something nobody can show yet.
The thing worth watching is outside testing. Maio points out that
TypeSafe hasn't said how Jev was built and trained, or how it checks
that its confidence numbers are honest. Tests by people with nothing
to sell, on real data that looks different from TypeSafe's, would
show whether those numbers hold up.
If your company's software makes thousands of small AI calls, ask
your team which of those calls are really decisions. Before anyone
switches tools, test on your own examples, because almost every
number in this launch comes from TypeSafe itself.
The bookshelf
I wrote the books below for people who build software. If you work
with a team building AI features, these are the books to hand them.
If you're starting to learn to build with AI yourself,
AI That Answers is the entry
point. It covers getting answers from AI models in a shape a program
can act on, which is the problem Jev is built around.
AI in TypeScript — five books, one path from your first LLM call to agents in production:
- AI That Answers — your first LLM app: prompts, structured output, token cost
- AI That Reads — RAG, embeddings, vector search, grounding in your docs
- AI That Acts — tool calling, functions, your first working agent
- AI That Plans — LangGraph.js, state, memory, multi-step and multi-agent
- AI That Ships — evals, guardrails, cost control, deploying on Node.js
Pocket Guides for Developers — short references you can finish in an evening:
- AI Agents Pocket Guide — patterns for building autonomous systems with LLMs
- LLM Observability Pocket Guide — picking tracing and evals tools for your team
- RAG Pocket Guide — retrieval, chunking and reranking for production
- Prompt Engineering Pocket Guide — techniques for getting the most from LLMs
Going deeper on tracing and evals: Observability for LLM Applications.
Top comments (0)