Skip the waitlist: running Jev's evaluate API via Vercel AI Gateway for $0.00074
AlchainHust · x · 2026-09-19
A hands-on walkthrough of using Jev's new evaluate API today via Vercel AI Gateway, no waitlist required.
Setup
- Create an API key in the Vercel console (AI Gateway → API Keys), store it as AIGATEWAYAPIKEY in .env; budget caps available. The author's 45 test calls cost just $0.00074.
- Use experimentalevaluate from the ai package with model: 'typesafe-ai/jev': state holds context, questions holds the classification tasks.
- Three question types: boolean / choice / score, returning typed answers plus per-option probabilities; there's no messages/chat interface.
Gotchas
- choice options are a map (each option needs criteria), not an array — arrays throw errors; score is the reverse, an array sorted low-to-high.
- confidence isn't in answers; it lives elsewhere in the response.
More from coding & agent
- Street Fighter poster animation on an 8GB GPU: MiniMax H3 + ComfyUI full breakdown — leftonredd33 · 2026-09-19
- AI coding agents' permission problem is becoming bigger than the model itself — Relevant-Potential17 · 2026-09-19
- Autonomous Labs open-sources OpenHarness, adds new senses to Lamp robot — dee_hw · 2026-09-19
- Six agent harnesses compared: token usage per task on identical model (Kimi K3) — zainhas · 2026-09-19
- Same model, 6 agent harnesses, 3x gap in end-to-end task completion time — zainhas · 2026-09-19
- First Official Hermes Agent Meetup Locks In Bangkok Date, Sponsored by Nous Research — Teknium · 2026-09-19