Meet Jev: TypeSafe's Classifier Model Built to Make Decisions — 200ms, Output Tokens Free

This isn't another chatbot. Jev is a new kind of model built to make structured decisions inside software — no string generation, nothing to hallucinate. A hands-on test against 1,000 real emails shows why that architecture changes the economics.

Video thumbnail — Meet Jev: The AI Built to Make Decisions
🎬 vogel ⏱️ 4:07 📅 Sep 2026
Classifier AI Structured Output System One Models TypeSafe

🧠 Jev: A Classifier, Not a Chatbot 0:00

The hook is a company origin story, and the video gets it slightly wrong. It says "the co-founder of ChatGPT" founded TypeSafe. The accurate version, straight from TypeSafe's launch post 🔗: the founder is Diogo Almeida, who worked at OpenAI on the methods that became the research behind ChatGPT — "the work that made language models useful at following instructions and talking with people." A key contributor to the ChatGPT research, not literally "co-founder of ChatGPT," and now the founder of TypeSafe.

The product is more interesting than the founder trivia. Jev is a classifier AI model — not a text generator. You feed it JSON rules, and it evaluates a set of rules against unstructured input: "based on this input, evaluate this set of rules." Production capability, churn risk, priority, spam — the kind of decisions software needs to make constantly. The headline claims: 20–200× faster and 40–400× cheaper, with output tokens free.

📌 Claim checked: TypeSafe's own launch post states the speed more conservatively as 40–200× faster than frontier LLMs "for the same levels of frontier intelligence" on structured-decision tasks, with input priced at $0.042 per million tokens (vs $0.20–$10 for frontier LLMs) and output tokens free. The video's "20–200× / 40–400×" is a looser paraphrase; the direction is accurate.

📬 The Email Test 0:54

Rather than read the marketing, vogel tests it on the most relatable workload there is: his own inbox. He exports 1,500 emails and gives Jev four things to evaluate for each — category, priority, spam, and reply. Those four are the core of email triage: "do you need to reply to it, or is it spam?" — the two questions that actually matter.

The setup is deliberately modest: a first batch of 100 emails with 8 parallel workers to measure the baseline, then a scale-up to 1,000. No fine-tuning, no special harness — just structured rules and the model doing classification.

⚡ Speed: ~200ms Per Email 1:17

The 100-email batch finished with numbers that made him do a double-take: ~200ms average per email, 240ms at the P95, and 38 emails per second. On the 1,000-email run it held steady — "it is flying, absolutely flying" — with 1.1 million input tokens processed and the per-email latency still hovering around 200ms.

Why that speed feels alien: "I'm not used to this type of speed of AI at all, because every other AI it's like — okay, well, we've got reasoning, we've got this and that." A frontier LLM reasoning for seconds per message is fine for chat, but a bottleneck inside a code loop. ~200ms full round-trip is the speed at which classification becomes a plain function call.

That last part is the important detail: the ~200ms is the full trip — email submitted, email returned — and it scales further with parallelism. This is why the architecture matters rather than just being fast.

🏷️ Classification Quality & Scale 1:34

Speed is meaningless if the calls are wrong, so he spot-checks the output. The priority flag correctly caught a security identifier in an OpenAI account email and a domain-revocation notice as high priority. The reply detector scored a real reply thread at 51%, and flagged an account-violation email at 91% — the ones that genuinely need a human to respond.

His verdict after paging through the classifications: "This did an amazing job… this is more than enough for standard email classification." The nuance he doesn't overclaim — this is a 4-minute hands-on demo, not a rigorous eval — but for the core triage decision (reply vs spam vs priority), the classifications landed where they should.

💸 The Cost & What Jev Actually Is 3:13

Then the cost, which is where he gets genuinely startled: after processing the full batch, the run had consumed about 5 cents. "This is insane… it's so incredibly cheap." The account's entire balance was $5, shared across the team, and the batch barely dented it.

What's actually under the hood, per TypeSafe's launch post, is a genuinely different class of model — what they call a System One Model:

Frontier LLMsJev (System One)
TrainingRLHF / RLVR (human preference, verifiable rewards)RLCD — Reinforcement Learning for Calibrated Decisions
OutputStrings — flexible, but must be parsed/validated, can hallucinateType-safe structured values, defined in advance, never type errors, calibrated probabilities
SamplingSequential — one token at a timeParallel — all outputs in a single query
Input cost$0.20–$10 / MTok$0.042 / MTok
Output cost~5× more than inputFree
Speed3–329s end-to-end70–500ms (40–200× faster)

The throughline is Almeida's own framing: "Think of Jev as a frontier-intelligence function call — unstructured state in, typed probabilistic decisions out." It gives up string generation, so it "can't hallucinate" in the sense that it never emits an invalid type. vogel's closing line captures what that means for builders: "Jev gives all AI properties of code."

💡 Key Takeaways

  1. Jev is a classifier, not a chatbot. It gives up string generation for typed, structured decisions that software can use directly — no parsing, no validation, no hallucinated types.
  2. The founder is Diogo Almeida, not a "ChatGPT co-founder." He worked at OpenAI on the research behind ChatGPT; the video's phrasing is imprecise.
  3. ~200ms per email at 38/sec. That's a full round-trip, holding steady from 100 to 1,000 emails — the speed where classification becomes a function call, not a reasoning pass.
  4. The cost is the other shock. The whole 1,000-email run cost about 5 cents; input is $0.042/MTok and output tokens are free.
  5. New stack, not a wrapper. Parallel sampling and RLCD (calibrated decisions) replace sequential token generation and RLHF — that's the source of the 40–200× speed gap.
  6. Built for automation, not conversation. TypeSafe's thesis: chat models have been superhuman for years, and the missing piece was fast, reliable structured decisions.

🔗 Resources & Links

Source video: youtube.com/watch?v=9oWxrsRo4d8

⏱️ Timestamp Index

0:00 A new type of AI: TypeSafe and Jev
0:54 Testing Jev on my own emails
1:17 First results: speed and latency
1:34 Reviewing the classifications
2:22 Scaling up to 1,000 emails
3:13 Checking the cost
☰ View all