Quick Rundown
-
Use Runway Gen-4.5 when you need precise shot execution.
-
Write your prompts like shot directions: Specify camera angle, movement, subject action, and timing.
-
Replace vague adjectives like "cinematic" or "dramatic" with visual instructions the model can execute.
-
Start with image-to-video for vertical output. Don't rely on text-to-video if you need 9:16.
-
Keep clips short and intentional. Plan around the 10-second limit before you generate.
-
Use the invideo Agent One to skip prompt engineering. Give it the goal, and let it structure the shot logic for you.
The cup that vanishes
Say your character is holding a cup.
They turn their head. The cup leaves the frame for half a second. The camera follows back, and the cup is in the wrong hand.
Or gone. Or holding a slightly different cup.
You run the same prompt a dozen times. Sometimes the cup survives. Sometimes it doesn't.
The failures cluster around the moments when the model has to remember something it can't currently see.
Runway's December 2025 release of Gen-4.5 is the first version where that failure mode reads as smaller.
Not solved. Smaller.
Independent reviews catalog three specific holes that still show up in complex scenes:
-
Objects that disappear after occlusion
-
Effects that arrive before their causes, like a door swinging open before the handle moves
-
"Success bias," where a sloppily aimed kick still finds the goal because the model has learned that kicks tend to score
Everything Runway shipped in Gen-4.5 makes sense once you accept what they were optimizing for.
What Runway shipped, and what they gave up
Read the Gen-4.5 specs as a list and they look incremental.
Read them as an argument and they're sharper.
The numbers:
-
2 to 10 seconds per generation
-
720p resolution
-
24 or 25 frames per second
-
Text-to-video locked to 16:9 (1280x720)
-
Image-to-video supports 16:9, 9:16, 1:1, 4:3, 3:4, and 21:9 ultrawide
-
12 credits per second on the Runway API
-
No native audio
-
Same speed and throughput as Gen-4
No native audio. The 10-second cap is half of what Seedance offers. The 720p ceiling is a step below the rest. Vertical text-to-video isn't supported out of the box; you have to start from a still and route through image-to-video to get a 9:16 result.
Every one of those decisions saved compute and capacity for one thing.
Runway tuned Gen-4.5 to follow detailed, sequenced camera direction inside a single prompt.
The three words that matter are detailed, sequenced, and inside.
Sora 2 reads camera language. Kling reads camera language. Seedance reads camera language.
None of them reads it the way Gen-4.5 does.
Gen 4.5 wins because, given the same prompt, it more often returns the shot the prompt described.
Which means the model's strengths only show up if the prompt actually describes a shot.
How to prompt Runway Gen 4.5
Here is a prompt that wastes Gen-4.5:
“Cinematic shot of a woman walking through a desert at sunset, beautiful lighting, photorealistic, dramatic mood.”
Every word in that prompt is asking the model to pick.
What is "cinematic" supposed to mean? Which lens? Which camera movement? How does she walk; toward the camera, past it, away from it? Where is the sun? Are we tracking, holding, pulling back?
The model has to answer all of those questions to render a single second of footage.
It answers them by averaging across whatever its training data associated with the words you used. Average is the visual equivalent of mush.
Here is a prompt that uses the model:
“Wide tracking shot, camera at hip height, moving slowly to the right alongside a woman walking across hard-packed desert sand toward a horizon lit gold by the last fifteen minutes of sunlight. Her shadow stretches long and to the left, slightly ahead of her stride. After three seconds the camera continues its lateral drift while she steps off-center to the left of frame; the shot widens into a static composition with her smaller, alone in the empty space. Hold for the final two seconds.”
The second prompt is longer. That isn't the point.
The second prompt names the lens height, the camera direction, the subject's gait against the surface, the lighting source and time, the shadow geometry, and a deliberate two-beat structure.
Every line is a thing the model can execute.
None of them is a feeling the model has to translate.
Image-to-video changes shape but not nature. When you start from a still, your first three sentences should describe what moves and how, not what the world looks like.
The world is already in the reference. The motion clauses do the work that the still can't.
Reference images for character consistency use the @ syntax that carried over from Gen-4. Tag the character, describe what they do, describe how the camera relates to them.
Motion Brush, also from Gen-4, lets you paint specific trajectory paths onto an input image. Useful when the model would otherwise default to a generic movement for a specific object.
The first kind of prompt asks it to imagine. The second kind asks it to execute.
How Runway fits inside the invideo workflow
Here is how you would use the Runway 4.5 inside invideo:
1. First, log in or sign up on invideo and start a new project. Then from the ‘Agent & Models’ select Runway.
2. Next, you draft a prompt of your choice. Make sure you follow the prompt advice we laid out before. Here, we will be using this prompt:
“Wide tracking shot, camera at hip height, moving slowly to the right alongside a woman walking across hard-packed desert sand toward a horizon lit gold by the last fifteen minutes of sunlight. Her shadow stretches long and to the left, slightly ahead of her stride. After three seconds the camera continues its lateral drift while she steps off-center to the left of frame; the shot widens into a static composition with her smaller, alone in the empty space. Hold for the final two seconds.”
3. Take the output and proceed with additional edits if required. This is the output we received against the prompt we fed Runway 4.5.
Agent One for the win
Invideo’s Agent One dramatically reduces the friction by removing the biggest bottleneck in AI video creation: prompt engineering. Instead of users needing to learn how to structure cinematic prompts, define camera motion, describe scene continuity, or manually translate an idea into model-friendly instructions, Agent One acts as an intelligent creative layer between the user’s intent and the underlying video model.
What makes this especially powerful is speed. Agent One makes intelligent decisions upfront, so users can move from idea to usable video far faster. For marketers, creators, and teams producing ads, social videos, explainer content, or branded campaigns, this translates into shorter production cycles and less dependence on specialized AI prompting skills.
Let’s get generating
The model isn't the limit anymore.
The limit is the language you bring to the prompt window.
Runway Gen-4.5 is the first version where this is true at the surface level, because it's the first version where the language you write is the language the model reads back.
Most people who try it will type the same description-of-a-feeling they've been typing into video models for two years and walk away convinced it's incremental.
A smaller number will sit with the change, write a real shot list, watch it execute, and realize that the bottleneck moved.
When the bottleneck moves, the question stops being which model.
It starts with whether you know how to direct. So get on invideo and start your journey today.
Frequently Asked Questions
-
1.
What is Runway Gen-4.5?
Runway Gen-4.5 is the latest video model from Runway, announced December 1, 2025, supporting text-to-video and image-to-video generation. It outputs 2 to 10 seconds at 720p, 24 or 25 fps, and currently holds the top position on the Artificial Analysis Video Arena leaderboard at 1,247 Elo. Its core differentiator is prompt adherence to detailed, sequenced camera direction inside a single prompt window.
-
2.
What's new in Runway Gen-4.5 compared to Runway Gen-4?
Runway Gen-4.5 improves on Gen-4 in three measurable ways. Motion quality, where objects carry more realistic weight and momentum and fluid dynamics are more natural. Multi-element composition, where occlusion, depth, and lighting hold up better when the scene has more than one or two subjects. Frame-to-frame consistency, where hair, fabric weave, and surface specularity drift less across motion. Same speed and throughput as Gen-4.
-
3.
Does Runway Gen-4.5 generate audio?
No. Runway Gen-4.5 outputs silent video. Audio has to be added in a separate pass, either via Veo 3.1 or Seedance 2.0 inside the same workspace, or through a traditional audio workflow.
-
4.
How long can a Runway Gen-4.5 clip be?
Runway Gen-4.5 generates 2 to 10 seconds per generation. Anything longer requires multiple generations stitched together. For longer single clips, Seedance 2.0 (up to 15 seconds) or Kling 3.0 multi-shot sequences are the alternatives.
-
5.
How much does Runway Gen-4.5 cost?
On the Runway API, Runway Gen-4.5 runs at 12 credits per second of generated video. On the Standard plan ($12 per user per month billed annually) the 625 monthly credits work out to roughly 25 seconds of Runway Gen-4.5 video. Some sources cite 25 credits per second for the web app at certain configurations; pricing schedules between Runway's public pricing page and dev docs were not perfectly aligned as of March 2026, so confirm on the specific plan you're on.
-
6.
Is Runway Gen-4.5 better than Sora 2 or Veo 3.1?
Better is the wrong frame. Runway Gen-4.5 wins blind preference benchmarks for shot-level visual fidelity and prompt adherence to camera direction. Veo 3.1 wins on native audio and lip sync. Kling 3.0 wins on 4K and multi-shot. Seedance 2.0 wins on multimodal references in a single pass. Pick by the job, not the leaderboard.
-
7.
Why does my Runway Gen-4.5 prompt produce generic-looking output?
Most likely because the prompt is describing a feeling rather than a shot. Runway Gen-4.5 reads camera direction (lens height, movement, timing), subject action (gait, gesture, beat structure), and lighting geometry. It does not reliably translate words like "cinematic," "dramatic," or "beautiful" into specific decisions. Replace adjectives with named camera moves, named lighting, and named subject actions in sequence.
-
8.
Does Runway Gen-4.5 support character consistency across shots?
Yes, through the @ reference system carried over from Runway Gen-4. Upload a reference image, tag it in the prompt, and the model will hold the character's appearance across generations. For dialogue with synchronized facial performance, Act-Two is the companion model that pairs with it.
-
9.
Can I use Runway Gen-4.5 for vertical video?
Runway Gen-4.5 image-to-video supports 9:16 vertical natively at 720x1280. Text-to-video is locked to 16:9 in the current release; for vertical text-to-video, generate a still image first and then route through image-to-video.
-
10.
Where does Runway Gen-4.5 still fail?
Object permanence in scenes with five or more discrete objects, where things vanish after occlusion. Causal sequencing in tight choreography, where effects can precede causes. Success bias, where actions tend to succeed even when the prompt asks for failure. Complex crowd or industrial scenes with many simultaneous moving elements.
-
11.
Is Runway Gen-4.5 worth it for short-form social content?
For shots where camera direction or motion fidelity matters and a single 10-second clip is the unit, yes. For dialogue-heavy talking-head content, Veo 3.1 is the stronger pick because of native lip sync. For multi-shot product walkthroughs, Kling 3.0 or Seedance 2.0 may fit better. Inside a workspace where all four are available, the answer is to use Runway Gen-4.5 for the shots that reward what it does.
-
12.
How do I access Runway Gen-4.5?
Runway Gen-4.5 is available across all paid Runway plans (Standard and higher), via the Runway web app, the Runway API, and inside invideo's model menu alongside Veo 3.1, Seedance 2.0, and Kling 3.0.
