You 0, Blitz 0, jev 0, luna 0.
Model shots are replayed from our experiments; you are the only live player.
Request API access
Blitz is in research preview. Leave your email and we'll get in touch about API access.
We use your email only to contact you about Blitz access.
The shooters
- Blitz — our ultra-low-latency decision model. Reacts in 22.5 ms (inference). Wrong shots: 6%. When it isn't sure, it holds fire and shows ? instead of guessing.
- jev — a hosted decision model. Reacts in 202 ms (end-to-end). Wrong shots: 18%. Never holds fire.
- luna — a frontier language model. Reacts in about 6.3 s (end-to-end). Wrong shots: 21%. Never holds fire; sits out rounds where it wasn't tested.
Wrong shots = confident answers that were wrong, across every text we tested.
Domains covered in this demo: tweet emotion, news topics, assistant requests, business workflows (security incidents, customer support, invoices, agent traces), coding-agent failure attribution.
How to play
- Read the card, then tap the bird or its card to shoot.
- +1 for a target, −1 for a wrong bird, 0 for holding fire.
- Two rounds are watch-only: the texts are too long to read in time.