Purpose
What Jev is and why it saves tokens
Jev is an AI that does not write. It makes decisions. You give it content and typed questions such as "Is this a lead?" or "How urgent is it, 1 to 5?", and it answers each one with a confidence score.
Jev was built by TypeSafe, a San Francisco startup founded by Diogo Almeida, a former OpenAI researcher who co-wrote the InstructGPT paper behind ChatGPT. It launched on 15 September 2026 and became the fastest-adopted model in Vercel AI Gateway history. TypeSafe says it is 40 to 200x faster and 40 to 400x cheaper than frontier models like Claude.
On hard judgment calls Jev trails Claude: 67.8% against 73.1% for Claude Opus 5 on TypeSafe's own benchmark. Use Jev for sorting and Claude for thinking.
Who this is for
- Founders and operators who run Claude on inboxes, DMs, comments or drafts
- Anyone who wants lower token spend without losing quality
Tools needed
- Claude
- An OpenRouter account
- $5 in OpenRouter credits
- A Mac with Terminal
No connector or MCP server is needed. The whole setup is an account, a key and one instruction to Claude.
Create an OpenRouter account
OpenRouter is the gateway that gives you Jev and hundreds of other models under one account.
- Open the link below in your browser
- Sign up with your email or a connected account
- Confirm your email if prompted
Add $5 in credits
- Open the Credits page in your OpenRouter account
- Add $5 using your preferred payment method
- Confirm the balance shows on the page
Sorting 50 emails cost $0.0013, so $5 lasts a long time. At that rate, 1,000 emails cost about 3 cents.
Create an API key
- Open the keys page below
- Click Create Key and give it a name you will recognise, such as Jev for Claude
- Copy the key. It starts with sk-or-
- Keep it open or in a safe place until you finish step 4
Treat the key like a password. It is billed against your credits.
Save the key on your Mac
Saving the key on your computer means Claude can use it without you pasting it into a chat.
- Open Terminal on your Mac
- Copy the command below and replace sk-or-... with your own key
- Paste it into Terminal and press Return
- Close Terminal and open it again so the new setting loads
echo 'export OPENROUTER_API_KEY="sk-or-..."' >> ~/.zshrc
Tell Claude to connect to Jev
- Open Claude
- Send this message exactly as written
- Claude writes a small helper for Jev
- The first time Claude sends your data to Jev, it asks permission. Click Allow
You only approve this once. From then on you can simply ask Claude to use Jev.
Run your first job
- Send Claude the message below
- Confirm each email comes back with a label and an urgency score from 1 to 5
- Check that the batch finishes in seconds, not minutes
- Open your OpenRouter activity page and confirm a small charge appears
If all four checks pass, the setup is complete.
Put it to work
Use cases that save Claude tokens
The pattern is always the same: Jev makes the quick call first, and Claude only runs on what is left.
| Use case | What Jev decides | What Claude skips |
|---|---|---|
| Inbox triage | Lead, client, team task, invoice or promo, plus urgency 1 to 5 | Reading every newsletter and notification (46% of our 50-email test) |
| Claude's bouncer | Does this request need Claude at all? | Every job that can be answered with a label |
| DM lead scoring | Buying intent 1 to 5 for each message | Reading cold DMs one by one |
| Brand checker | Does this caption follow my brand rules? Yes or no | Full rewrite passes on drafts that already pass |
| Comment sorting | Question, objection, testimonial or spam | Scanning hundreds of comments for content ideas |
| Support routing | Which FAQ or template fits this ticket | Writing replies to the same five questions |
| Content QA | Is this hook strong enough, 1 to 5? | Rewriting hooks that score 4 or 5 |
| Safe auto-approve | Is this action low-risk? | Back-and-forth permission checks |
Start with inbox triage. It gives the clearest before and after, and the 46% of emails Jev filters out never reach Claude.
Reference
Technical reference for Claude
Claude writes this helper for you in step 5. Keep these values for debugging or custom builds.
| Setting | Value |
|---|---|
| Endpoint | POST https://openrouter.ai/api/alpha/decisions |
| Model | ~typesafe/jev-latest |
| Question types | choice, noul (yes/no) and score |
| Key variable | OPENROUTER_API_KEY |
FAQ
Common questions
Does Jev replace Claude?
No. Jev only decides: yes or no, a label, or a score. It cannot write. Claude still handles writing and hard judgment calls.
How much does Jev cost?
$0.042 per million input tokens. Its answers are free. Our 50-email test cost $0.0013.
Is Jev accurate enough?
For sorting and scoring, yes. On hard judgment calls it scores 67.8% against 73.1% for Claude Opus 5 on TypeSafe's own benchmark, so route those to Claude.
Do I need an MCP server or connector?
No. You need an OpenRouter account, a saved key and one instruction to Claude.
Can I use other models through the same account?
Yes. One OpenRouter account gives you Jev and hundreds of other models.
Troubleshooting
If something stalls
| Problem | Fix |
|---|---|
| Claude says it cannot find the key | Close Terminal, open a new window and ask Claude to connect again. Confirm the command in step 4 ran without errors. |
| Requests fail or are refused | Check your balance on the OpenRouter Credits page and top up if it is empty. |
| Claude asks permission every time | Click Allow when prompted. Approval is requested the first time Claude sends your data to Jev. |
| Key was pasted incorrectly | Create a new key at openrouter.ai/keys, then repeat step 4 with the new value. |
Sources
Where the facts come from
- Jev on OpenRouter (docs)
- DataCamp: System One models and Jev
- innFactory: Jev is a classifier, not an LLM
- The Next Web: The fastest-adopted model on Vercel this year cannot write a word
Speed, cost and accuracy figures are TypeSafe's own claims. The inbox numbers are from our own test on 5 October 2026.
Version history: v1.0, 7 October 2026, first release.