The Lab
Four hands-on stages in this browser. Make a real model call, extend a rule-based till until it breaks, write your own prompt, then score it against twenty real cases and watch the number move.
First, connect a model
Three of the four steps below send your words to a real AI model and show you what comes back. For that the Lab needs a key of your own.
It is a long password that lets a program talk to an AI model on your behalf. You create one on the model provider's website, and it belongs to you, not to us. You pay the provider for real requests; the exact amount depends on model, time, cache use and returned usage.
- Open an account with DeepSeekSigning up needs a phone number, and a live call needs a small top-up. DeepSeek's page shows the minimum; add only an amount you are comfortable spending. The Lab shows request caps and an estimate before each paid action. If you cannot, open the last step and choose “No key? Show a scripted run instead”.Get a key from DeepSeek
- Create a key and copy itCopy the complete key. DeepSeek may show it only once; this Lab does not reject a key merely because of its prefix.
- Paste it in the box belowIt stays in this browser tab and is erased the moment you close it.
🔒 Held in session storage for this browser tab and erased when the tab closes. Current site code sends it only to DeepSeek. Vercel Analytics runs on this same origin and is therefore a third-party trust boundary; current site code does not send it your key, prompts or replies.
Recommended journey: about $0.04 in total — up to 60 paid requests and 16,350 output tokens (1 + 3 + 28 + 28). Each extra evaluation can add up to 28 requests. Batches use concurrency 4.
Pre-run estimates use peak price, cache misses and the maximum output caps. Actual totals use returned usage. Price snapshot checked 2026-09-22. DeepSeek pricing source
Stop prevents new requests and attempts to abort at most four already in flight. A request already sent may still be billed; without returned usage its cost stays unknown.
Your first call
Before anything clever: prove the wire works. Ask the model anything, and watch what it costs — almost every tutorial hides that number.
Run sends 1 paid request to deepseek-flash with a 250-token output cap. Peak/cache-miss estimate: up to about $0.00034; actual cost uses returned usage.