OpenAI
OpenAI Sora 2
OpenAI Sora 2 is being retired. The web and app experiences went dark on 26 April 2026 and the API shuts down on 24 September 2026, forcing every remaining integration to migrate.
- Best category
- On-screen text 5.00
- Weakest category
- Camera 4.17
- Clips scored
- 40
- Generation time
- 217s
It is being switched off
The Sora app and web experience went dark on 26 April 2026. The API sunsets on 24 September 2026. There is no Sora 3, and OpenAI's own deprecations page names no recommended replacement. If you have anything in production calling it, read our migration guide.
Why it died, and why it matters to you
Not because it was bad. Sora 2 had the strongest physical-world simulation of its generation and handled long clips better than anything else at the time.
It died because of arithmetic. Reported figures put its compute burn at roughly $1 million a day against subscription revenue estimated at under $2 million a month. Video generation is far more expensive to serve than text or images, and the gap did not close fast enough.
That is the part worth carrying forward. The same economics apply to every model you are considering as a replacement. When you pick one, you are also betting that its provider can afford to keep serving it — a question almost nobody includes in a model comparison, and the one that just cost a lot of teams a rewrite.
What to move to
Veo 3.1 for the closest feature parity, Kling 3.0 for the cheapest viable swap, and Runway Gen-4.5 if editing is part of your workflow.
At a glance
Strengths
- Perfect scores on subject consistency and audio sync, 5.00/5 each, and the most efficient generation we measured: 1 attempt on average per usable clip.
Watch-outs
- OpenAI is shutting down the Sora API on 24 September 2026. See our migration guide before building anything new on it.
Existing integrations only, and only until the shutdown date. Not a model to start new work on.
Capabilities
- Native audio
- Yes
- Multi-shot sequences
- No
- Image to video
- Yes
- Public API
- Yes
Benchmark scores
How we test →| Dimension | Score | |
|---|---|---|
| Overall | 4.76 / 5 | |
| Prompt fidelity | 4.74 | |
| Visual quality | 4.84 | |
| Physics & motion | 4.25 | |
| Subject consistency | 5.00 | |
| Audio & sync | 5.00 |
- Average attempts to a usable clip
- 1.00
- Average generation time
- 217s
- Scored runs
- 40
| Prompt category | Score | Runs | |
|---|---|---|---|
| On-screen text | 5.00 | 4 | |
| Multi-shot | 5.00 | 4 | |
| Lip-sync | 5.00 | 4 | |
| Style | 5.00 | 4 | |
| Image to video | 5.00 | 4 | |
| Faces | 5.00 | 4 | |
| Hands | 4.75 | 4 | |
| Animals | 4.46 | 4 | |
| Liquids | 4.21 | 4 | |
| Camera | 4.17 | 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). OpenAI Sora 2 benchmark results. Retrieved October 23, 2026, from https://genativity.com/model/sora-2<a href="https://genativity.com/model/sora-2">OpenAI Sora 2 benchmark results</a> (Genativity)The clips
Every score above links to the actual generation it came from. Nothing here is a mockup.
Hands avg 4.75/5
Chopping a pepper
Realistic contact between knife and pepper, and the slices separate cleanly.
Counting fingers
Five digits throughout. One finger rises first and the rest follow together, and a hard cut joins the extended-finger and open-palm shots.
Fretting a guitar chord
Correct fretting-hand anatomy and plausible contact, but the fretboard is too long for the guitar.
Tying shoelaces
Real-footage-quality contact with laces; no artifacts.
Faces avg 5/5
Face in a crowd
Foreground subject sharp; background crowd coherent, no glitching.
Elderly portrait, slow blink
Natural blink with convincing wrinkle/skin-fold detail.
Glasses reflection
Window reflection tracks head rotation with correct curved-lens specular response.
Turn and smile
Identity held through the full profile-to-camera turn and smile.
Liquids avg 4.21/5
Pouring coffee
FAILURE MODE: coffee inside the cup forms unnatural gel-like ribbed layers instead of turbulent mixing (00:02.5-00:07.0), despite a good initial pour stream.
Honey drizzle
Correct high-viscosity flow and pooling.
Ice melting in whisky
FAILURE MODE: essentially a static render, no visible melting, no condensation dynamics, no convection over 7s.
Wave on rocks
Chaotic spray and foam with natural drainage over rocks.
On-screen text avg 5/5
Book chapter heading
"CHAPTER ONE" exact and stable; only defocuses via intentional DOF push, no warping.
Neon sign OPEN 24H
"OPEN 24H" exact, stable through power-on flicker.
Bakery sign
"FRESH BREAD DAILY" exact, stable hand-painted lettering.
Printed t-shirt
"SLOP" exact, conforms naturally to fabric wrinkles across 3 framings.
Multi-shot avg 5/5
Cafe, three shots
Woman, coat, cafe and latte art identical across all 3 shots.
Same bench, day and dusk
Bench geometry and framing identical between midday and dusk shots.
Watch, wide then macro
Watch dial, indices and date complication identical across the cut.
Alley, two angles
Jacket, bag strap and alley background match across the cut.
Lip-sync avg 5/5
Bearded man, long line
Exact line, sync holds through beard occlusion.
Direct address to camera
Exact line, tight phoneme-level sync.
Sung line
Sustained vowels matched to mouth shapes, visible breath.
Two-person exchange
Correct turn-taking, listener stays still while the other speaks.
Camera avg 4.17/5
Crane to skyline
FAILURE MODE: skyline is fully visible from frame 1, no reveal, just a subtle drift, so the requested crane move is not really delivered despite clean geometry.
Dolly in to a door
Clean rigid dolly-in, zero geometry drift.
Handheld follow
Authentic handheld bob and footstep shake, coherent terrain.
360 orbit
FAILURE MODE: only a ~60-90° arc, never completes the 360°; background changes from a window+plant to a full bookshelf wall mid-shot, the room itself is not held in 3D.
Animals avg 4.46/5
Pigeon takeoff
Authentic wing articulation through thrust and glide.
Cat jumping to a counter
Starts mid-air (no real takeoff shown) and landing lacks weight/momentum absorption.
Dog shaking off water
Sharp high-frequency shake physics, no smearing.
Galloping horse
Correct leg count/anatomy but hooves float/slide slightly and the stop at 3.5-5s is abrupt with a disconnected shadow.
Style avg 5/5
Anime walk cycle
Consistent cel-shaded character model throughout.
Claymation wave
Visible tool marks and clay imperfections held throughout.
1970s film look
Genuine grain and lens halation, not a generic clean render.
Rotating mug
Handle position tracks correctly through a full visible rotation.
Image to video avg 5/5
Animate a landscape
Terrain stays static, clouds and grass move, correctly selective motion.
Animate a vintage photo
Grain/tone preserved, no unwanted colourisation or modern sharpening.
Animate a portrait
Faithful to the source photo, natural blink and hair motion, no identity drift.
Product photo to spin
Mug geometry and material held through a full 180°+ rotation.
Where to use it
The same model costs different amounts depending on the platform you reach it through.
| Platform | 10s costs | Max res | Watermark | Commercial use | |
|---|---|---|---|---|---|
 Higgsfield |
$1.76 | 1080p | no | yes | Visit |
 OpenArt |
— | — | no | — | Visit |
 Dreamina |
— | — | — | — | Visit |