Live50 new in 24h/Ollama: Ollama v0.34.4
aimluptodate
My Stack
Preview · "(demo)" items are samples. How we verify

← Claim Check

The claim · 24 Sept 2026

Jev is 193x faster and 445x cheaper than LLMs.

Seen on: "LLMs vs. Jev, clearly explained" (X, widely reshared on LinkedIn) ↗

True, with a catch

Jev really is far faster and cheaper, but only for yes/no, pick-one and score-it decisions. It can't write text.

The 193.6x / 444.6x figures are TypeSafe AI's own numbers from its workflow tests against GPT-6 Astra and Fable 5.1. An early independent test by OpenRouter shows the same pattern (about 4–6x faster than GPT Luna at the same accuracy) on classification tasks. Jev doesn't replace an LLM: it can't generate text, so it only handles decisions.

Disagree / report an error

The numbers

Speed vs LLMsup to 193.6xTypeSafe's own tests
Vendor-documented
Cost vs LLMsup to 444.6xvs GPT-6 Astra, Fable 5.1
Vendor-documented
Latency70–500 msend to end
Vendor-documented
Input price$0.042per 1M tokens · output free
Vendor-documented
Ticket triage, median194 msGPT Luna 1,106 ms · Claude Opus 1,957 ms
Community-reported
Options per question≤ 255defined up front
Vendor-documented

What's true

  • Jev is a new kind of model from the startup TypeSafe AI, launched in early access on September 15, 2026. It is not an LLM: it doesn't write text. You give it some state (a support ticket, an agent's progress) and questions, and it returns typed answers with probabilities. Vendor-documented
  • It answers three kinds of question: Boolean (probability that something is true), Choice (pick one option from a list you define) and Score (rate against a scale you define). All questions are answered in parallel in a single pass. Vendor-documented
  • Responses take 70–500 ms. Input costs $0.042 per million tokens and output is free. Vendor-documented
  • It's available through Vercel AI Gateway (model id typesafe-ai/jev) via the AI SDK, an HTTP API or TypeSafe's own clients. Within 24 hours, nearly 13% of paid AI Gateway teams were using it, Vercel's fastest model adoption ever. Vendor-documented
  • In OpenRouter's test of routing 60 support tickets, Jev matched GPT Luna's accuracy (98.3% on intent) at 194 ms vs 1,106 ms median latency, for about a quarter of the cost. Claude Opus was slightly more accurate on intent (100%) but took 1,957 ms and cost over 100x more. Community-reported

The catch

  • The headline 193.6x and 444.6x are best cases from TypeSafe's own workflow tests; the company itself says they are 'on the higher end' of real-world gains. In OpenRouter's test it was about 4–6x faster than a fast LLM. Vendor-documented
  • Jev gives up text generation completely. It can't summarise, explain, write code or answer open questions, so it complements an LLM rather than replacing one. Vendor-documented
  • Every answer must come from options you define up front, at most 255 per question. It works on text/structured state, not images. Vendor-documented
  • It's in early access from a brand-new company. Independent benchmarks so far are small (60 tickets, 40 messages), and pricing and limits may change. Community-reported

What the post got wrong

  • It's "Jev" from TypeSafe AI, a "System One" decision model. It's sometimes mixed up with JEPA, Yann LeCun's unrelated world-model research.
  • "445x cheaper than LLMs" compares against large frontier models on decision tasks. Against a small, cheap LLM the gap is much smaller (about 3–4x cheaper than GPT Luna in OpenRouter's test).

How it works, in plain words

An LLM answers by writing, one token at a time, and your code then parses the text. Jev skips the writing: it reads the whole input once and scores every option you allowed, like a multiple-choice exam instead of an essay. Fewer steps means much lower latency and cost, and you get a real probability you can branch on (auto-approve above 0.9, send to a human below). The suggested pattern: route with Jev, compute in code, write with an LLM.

Worth trying if you are…

  • Teams routing support tickets, emails or requests into fixed queues
  • Agent builders deciding the next tool, whether to retry, or when to stop
  • Guardrail and policy checks (prompt-injection screening, 'is this safe to send?') and anything that ends in a switch statement

Skip it if you need…

  • Anyone who needs written output: chat, summaries, code, explanations. Keep your LLM for that
  • Open-ended questions where you can't list the possible answers in advance
  • Image, audio or video inputs

Try it through Vercel AI Gateway

  1. Create a Vercel account and an AI Gateway API key.
  2. Install the AI SDK 7.0.105 or later, which adds the experimental evaluate API.
  3. Call the model typesafe-ai/jev with your state and named questions (Boolean, Choice or Score). From other languages, POST to AI Gateway's /v1/evaluate endpoint.
  4. Branch on the returned probabilities: automate confident answers and send uncertain ones to a person.
npm install ai@latest   # 7.0.105+ has the experimental evaluate API

# model id: typesafe-ai/jev   (billed through Vercel AI Gateway)
# question types: boolean | choice | score
# full examples: vercel.com/changelog/typesafe-ai-jev-now-available-on-ai-gateway

Sources

Our own test

Scheduled: Jev vs a small LLM vs a frontier LLM on our ticket-routing and guardrail suites (500+ items), measuring accuracy, calibration of the probabilities, latency and cost per 1,000.

Want something that works today?

Cut my AI API bill

Most teams can cut spend substantially without losing quality by applying four levers in order.

Open the guide