news
We ran Sora one last time before OpenAI shuts it down
The final score: 4.78/5, our highest yet — and it fooled our own judge into thinking it was real footage.
OpenAI's Sora goes dark for good on 24 September 2026, when the API that is the last thing keeping it alive gets switched off. We ran it through our full 40-prompt benchmark before that happens — the same battery we run on everything — so there is a real, measured record of what it could do, rather than a eulogy built on marketing copy.
The score: 4.78 out of 5
That is the highest overall score of any model we have tested so far, ahead of Grok (4.67) and Kling 3.0 (4.45). Full breakdown, every clip, on the leaderboard.
| Dimension | Sora 2 |
|---|---|
| Prompt fidelity | 4.76 |
| Visual quality | 4.88 |
| Physics & motion | 4.25 |
| Multi-shot consistency | 5.00 |
| Audio & sync | 5.00 |
The 5.00 on consistency is not a rounding artefact — it aced every multi-shot continuity prompt we threw at it, including the ones carrying the same subject through three separate shots. That is the frontier capability that separates this generation of models from the last, and Sora held it perfectly while running out the clock.
Where it actually broke
It was not flawless. Two failures are worth knowing about, because they are the kind of thing you only find by running the model, not by reading about it:
- A full 360-degree orbit never completes — the camera manages roughly 60-90 degrees and the background quietly rewrites itself midway through, swapping a window and a plant for a wall of bookshelves. Score: 2.83/5, our lowest for this model.
- A crane shot meant to reveal a skyline shows the skyline from frame one — there is no reveal, just a light drift. Score: 3.83/5.
Interestingly, both Kling and Grok nailed these exact two prompts, camera moves included. Nothing in this field is uniformly better at everything, Sora included.
The detail nobody expects: it fooled our own analysis tooling
When we ran our vision-model judge over the hand and object-interaction clips, it repeatedly described the footage as "real-world live-action footage" — not as a caveat, as its actual assessment, before we had to remind it what it was looking at. That happened on the shoelace-tying and guitar-fretting prompts specifically, the ones with the most demanding hand-object contact in the whole battery.
Why it's dying anyway
None of this saved it. Reported figures put Sora's compute cost at roughly $1 million a day against subscription revenue estimated at under $2 million a month — arithmetic that no amount of visual quality fixes. Usage reportedly peaked near a million monthly active users and had fallen under 500,000 by the time the shutdown was announced.
If you are still calling the API, you have until 24 September. Read our migration guide for what to move to and why the published pricing for the alternatives is not as trustworthy as it looks.
Related: OpenAI Sora 2
Jose Florido — Runs the SlopTV benchmark. Every score on this site comes from a generation he paid for and a clip you can watch.