While the world was busy arguing about Gemini 3 and GPT 5.2, a smaller player quietly stole the crown. Runway has officially released Gen 4.5, a new video generation model that currently sits at #1 on the global leaderboards.
Internal documents reveal the model was codenamed “David” during development, a deliberate nod to the startup’s battle against tech giants (Goliaths) like Google and OpenAI. And if the latest benchmarks are any indication, the sling just hit its mark.
The “Physics” Update We Were Waiting For
The biggest complaint with previous AI video models (including Gen 3 Alpha) was “floatiness.” Characters glided rather than walked; objects lacked weight. Gen 4.5 specifically targets Physical Accuracy.
According to both the official documentation and early user tests, the model now understands:
- Momentum and Weight: Heavy objects accelerate and stop realistically. They do not just drift.
- Liquid Dynamics: Water, rain, and spills now follow correct fluid physics rather than morphing into weird sludge.
- Complex Interactions: Multi object scenes where items bounce or collide (like the “glass sphere on marble staircase” test) now maintain coherence without clipping through each other.
Also Read: GPT 5.2 Review: Why the Backlash Despite Record-Breaking Benchmarks?
Benchmarks: Gen 4.5 vs The Giants
The most shocking data comes from the independent Video Arena leaderboard (maintained by Artificial Analysis). Runway has overtaken both Google’s latest studio model and OpenAI’s flagship pro model.
| Model | Elo Score | Ranking |
|---|---|---|
| Runway Gen 4.5 | 1247 | #1 |
| Google Veo 3 | 1226 | #2 |
| OpenAI Sora 2 Pro | 1206 | #3 |
| Kling 1.6 | 1180 | #4 |
Note: The Elo score represents blind human preference, meaning users consistently picked Runway’s video quality over the others in side by side tests.
What is New? Key Features
- Prompt Adherence: The model is significantly better at following complex, multi clause instructions. If you ask for a specific camera move and a specific character emotion, it is less likely to ignore one of them.
- Emotional Consistency: Characters can now display subtle “micro expressions” (like holding back tears or a slight frown) without their faces distorting or morphing into someone else.
- NVIDIA Powered: The model was trained entirely on NVIDIA’s new Blackwell GPU cluster, which likely contributes to its speed and inference efficiency.
Pricing and Availability
Unlike Sora 2 (which is still gatekept behind high tier subscriptions) or Veo (limited access), Runway Gen 4.5 is rolling out to all users immediately.
- Free Plan: Includes limited credits (approx. 105s of generation) to test the model.
- Standard Plan ($12/mo): 625 credits/month.
- Pro Plan ($28/mo): 2250 credits/month (approx. 90 seconds of Gen 4.5 video or significantly more of Gen 4 Turbo).
- Unlimited Plan ($76/mo): Unlimited generations (relaxed rate mode).
The “Gap”: What No One Is Saying
While the physics update is massive, early stress tests show the “Causality Problem” isn’t 100% solved. In extremely long shots (over 10 seconds), objects can still occasionally “forget” they exist if they leave the frame and come back. However, for the 5 to 10 second clips most creators use for B-roll or social media, it is currently the most stable model on the market.
Verdict: If you need realistic B-roll where gravity actually works, Runway is currently beating the trillion dollar giants at their own game.
