Finish the next clip before playback ends
Next-scene generation
Compare complete-clip generation with RTR, then check whether the published number covers inference or the full request.
API COMPARISON · 7 DOCUMENTED
Real-time AI video APIs, compared by the product loop.
A fast clip endpoint, a WebRTC transformation stream, and an interactive world API solve different problems. Compare access and evidence without flattening them into one speed rank.
| Model / source | Realtime class | Transport | Native audio | Published timing | Price | Evidence |
|---|---|---|---|---|---|---|
| H3 Maxfalfal launch announcement | CLASS A · FASTER CLIP | Queued API | Yes | RTR 0.60 (inference) | $0.08 / output second at 768p | OFFICIAL CLAIM |
| H3 Max Turbofalfal API documentation | CLASS A · FASTER CLIP | Queued API | Not reported | Not reported | Not reported | OFFICIAL CLAIM |
| HeliosReactorReactor model catalog | CLASS B · STREAMING | WebRTC | Not reported | Not reported | Not reported | OFFICIAL CLAIM |
| Lucy 2.0DecartDecart official site | CLASS B · STREAMING | WebRTC | No | Not reported | Not reported | OFFICIAL CLAIM |
| MiniMax H3MiniMaxMiniMax open-source announcement | TRACKED | API / local deployment | Yes | Not reported | Not reported | OFFICIAL CLAIM |
| MirageDecartMirage official demo | CLASS B · STREAMING | Live stream | No | Not reported | Not reported | OFFICIAL CLAIM |
| Oasis 2.0DecartOasis 2.0 official experience | CLASS C · INTERACTIVE WORLD | Interactive stream | Not reported | Not reported | Not reported | OFFICIAL CLAIM |
“Not reported” means the current source does not provide a comparable value. It never means zero cost or zero latency.
CHOOSE BY JOB
Three realtime classes. Three buying questions.
Change an active visual stream
Live transformation
Transport, time to first frame, cadence, jitter, and mid-stream control matter more than clip RTR.
Let actions alter future generated frames
Interactive worlds
Look for action-to-effect latency, output FPS, supported controls, persistence, and session limits.
BEFORE YOU INTEGRATE
Ask where the clock starts and stops.
Official inference time can be useful for capacity planning. End-to-end waiting also includes queueing, preprocessing, encoding, storage, and delivery. Keep both values and their conditions separate.