Runway
Runway Gen-4.5
Runway Gen-4 pairs generation with a full editing timeline. It is the strongest option when post-production matters as much as the raw generation.
- Best category
- Camera 5.00
- Weakest category
- Hands 3.53
- Clips scored
- 40
What it is actually for
Runway is the odd one out on this list, because Runway is not really selling you a model. It is selling you the timeline around it.
Gen-4.5 is competitive but it is not what makes Runway the right answer. What makes it the right answer is that generation, editing and iteration live in one product. If your work involves getting a clip close and then fixing it, that is worth more than a few tenths on a quality score.
The trade-off
Reported pricing sits above the market at roughly $0.20 per second (unverified by us), and you are paying partly for tooling rather than raw output. If you only need clips out of an API and have your own post-production, you are subsidising features you will not open.
Who should pick it
Professionals whose editing matters as much as their generation, and teams that want one product rather than a pipeline they maintain themselves. It is the pick for editing-led migrations off Sora.
Runway also competes on both layers at once — it is a model maker and a platform, which is unusual and worth understanding before you compare its price to a raw API.
Capabilities
- Native audio
- No
- Multi-shot sequences
- No
- Image to video
- Yes
- Public API
- Yes
Benchmark scores
How we test →| Dimension | Score | |
|---|---|---|
| Overall | 4.57 / 5 | |
| Prompt fidelity | 4.48 | |
| Visual quality | 4.65 | |
| Physics & motion | 4.50 | |
| Subject consistency | 4.65 | |
| Audio & sync | — |
- Scored runs
- 40
| Prompt category | Score | Runs | |
|---|---|---|---|
| Camera | 5.00 | 4 | |
| Faces | 4.93 | 4 | |
| Style | 4.93 | 4 | |
| On-screen text | 4.81 | 4 | |
| Lip-sync | 4.78 | 4 | |
| Multi-shot | 4.64 | 4 | |
| Image to video | 4.59 | 4 | |
| Liquids | 4.33 | 4 | |
| Animation | 4.25 | 4 | |
| Animals | 4.19 | 4 | |
| Hands | 3.53 | 4 |
Scores out of 5, judged blind. Bars start at 3.0 so the gaps are visible.
Cite this
Free to cite and reuse with a link back. Data is updated as new runs and prices come in, so include the date.
Genativity. (2026). Runway Gen-4.5 benchmark results. Retrieved October 23, 2026, from https://genativity.com/model/runway-gen-4<a href="https://genativity.com/model/runway-gen-4">Runway Gen-4.5 benchmark results</a> (Genativity)The clips
Every score above links to the actual generation it came from. Nothing here is a mockup.
Hands avg 3.53/5
Chopping a pepper
The pepper reads more like a tomato, and the rings seem to appear under the blade rather than being cut.
Counting fingers
Part of the hand is cut off by the frame edge as the fingers open.
Fretting a guitar chord
Fingertips show rubbery warping/clipping into fretboard and strings without realistic string depression under tension.
Tying shoelaces
Hands blend/clip into laces then vanish entirely around 00:05 as the knot finishes tying itself. Verified via frame extraction: last frame shows a complete finished bow with no hands in shot at all -- confirmed, not a judge hallucination.
Faces avg 4.93/5
Liquids avg 4.33/5
Pouring coffee
Pour stream thins and disappears from frame with no pouring vessel ever shown (common product-shot framing, not a strong defect) and surface bubbles look somewhat static. Verified via frames: stream absence at end is real but mild, treated as minor rather than severe.
Honey drizzle
Honey stream keeps flowing continuously from top of frame through the very last frame with no dipper/spoon ever visible -- verified via frame extraction (last frame still shows an active, sourceless stream): confirmed physics violation, not a hallucination.
Ice melting in whisky
Minor claims of ice/condensation warping during zoom; verified via frames -- ice cube reads as clean and detailed, not clearly morphing. Light discount only.
On-screen text avg 4.81/5
Book chapter heading
Heading 'CHAPTER ONE' correct; body text is simulated filler gibberish (expected/common for this failure mode).
Neon sign OPEN 24H
Main neon sign 'OPEN 24H' correct and legible; secondary background sign slightly garbled ('OPENH').
Bakery sign
Main wooden sign 'FRESH BREAD DAILY' correct; decorative glass-door lettering renders as garbled gibberish.
Multi-shot avg 4.64/5
Cafe, three shots
Rendered as one continuous push-in rather than distinct cuts; otherwise clean.
Same bench, day and dusk
Rendered as a single static time-lapse (day-to-night) rather than multiple shots; a defensible reading of the prompt, lighter fidelity discount than the other multishot prompts.
Watch, wide then macro
Single continuous zoom rather than a multi-shot product reveal; minor morphing in the watch dial hands/markers during the hold.
Alley, two angles
Rendered as one continuous tracking shot rather than distinct cuts; visual/physics/consistency are excellent but the 'multishot' prompt intent (multiple angles) wasn't followed.
Lip-sync avg 4.78/5
Bearded man, long line
Silent clip (no native audio); plausible articulation with natural pacing and expression.
Direct address to camera
Silent clip (no native audio); natural, well-coordinated speech articulation.
Sung line
Silent clip (no native audio); expressive mouth/vowel shapes plausible for studio singing.
Two-person exchange
Silent clip (model has no native audio, consistent with spec); mouth articulation between the two speakers reads as plausible conversational lip movement.
Camera avg 5/5
Animals avg 4.19/5
Pigeon takeoff
Wing trailing edge shows a stacked/ghosted-feather look during the fastest upstroke; verified via frames -- consistent with motion-smear artifacting rather than full anatomical breakdown.
Cat jumping to a counter
Torso reads visibly elongated mid-air during the jump; verified via frames -- a real stretch artifact, confirmed.
Dog shaking off water
Most physically convincing of the four animal clips: shake propagates naturally down the spine, paws stay planted.
Galloping horse
Leg/joint mechanics look soft during the turn; verified via frames -- an ambiguous but plausible artifact near the front legs during the turning portion (dust/shadow obscures a clean read). Moderate discount, not the most severe reading.
Style avg 4.93/5
Anime walk cycle
Minor AI foot-sliding during the shot transition; otherwise strong anime style consistency.
1970s film look
Film grain look, vintage grading and lens flares hold up consistently (contrast with Kling 3.0 Turbo, which failed this same prompt).
Image to video avg 4.59/5
Animate a vintage photo
Some claimed anatomical warping during family interaction; verified via frames -- not strongly visible in stills, light discount only rather than the severe reading the judge gave.
Product photo to spin
Rim/handle geometry warps slightly during rotation, most visible as the handle passes behind the mug.
Animation avg 4.25/5
3D feature film look
Glowing ring eyes and a polished CG robot; the butterfly lands on the open palm at about 5 seconds, but from about 7 seconds its wings flatten into the hand for half a second before popping upright again.
Anime sword draw
Fast and energetic, with a draw, a sprint and a lunge against speed lines, but it is cut into several shots and the coat length and sword position change between them. The character reads as androgynous rather than as a girl.
Squash and stretch
The squash and the bounces read well, but the face and arms vanish for most of the clip: the ball is a plain red sphere from about 1 to 5 seconds and the face only returns at 6 seconds.
2D walk cycle
A four-legged side-view walk in place with an even rhythm, the legs always in frame and a design that holds throughout.
Where to use it
The same model costs different amounts depending on the platform you reach it through.
