The first Sora launched as a preview in February 2024 and then went quiet. For eighteen months it was the model everyone had heard of and almost no one could use. Sora 2 ended that. Released on September 30, 2025, paired with a consumer app, and shipping with the Cameos feature that immediately broke the App Store charts, it is now genuinely available. That changes the conversation about what an AI video model is for.
I have spent the last three weeks running Sora 2 through the kinds of prompts we ship for client work and through the kinds of dumb-fun prompts the Sora app was clearly built for. The headline: it is a remarkable model trapped inside an app-first product. The model is excellent. The way OpenAI ships it forces you to decide whether you are buying a creative tool or a TikTok competitor.
What is Sora 2 (and how it differs from Sora 1)
Sora 2 is OpenAI’s second-generation video model. It generates 1080p video with native synchronized audio, supports multi-shot scenes, and is tuned for physics-aware motion. Three concrete deltas vs the 2024 preview model: audio is now generated in the same pass as the video; physics simulations (water, fabric, collision) are dramatically more reliable; and a new feature, Cameos, lets you insert a verified likeness of yourself into any prompted scene.
The model is distributed through the Sora app (iOS and web), through ChatGPT (on the Pro plan), and through the OpenAI API in a tiered rollout. Practical access depends on which plan you pay for.
How I got access
I downloaded the Sora app, signed in with my OpenAI account, and got daily-quota access immediately on a free tier. ChatGPT Plus at $20/mo widened the quota and unlocked longer clips. To run the high-tier tests in this review I upgraded to ChatGPT Pro at $200/mo for one month. API access is still tiered; my organisation got audio-enabled API last week after ID verification.
The four prompts I used
Same methodology as our Veo 3 review, so you can compare directly. Each prompt targets a different model strength.
- Cameo with dialogue. My uploaded cameo delivering a 6-word line in a coffee shop, with steam from the espresso machine.
- Physics-heavy action. A wave breaking against a rocky pier at sunset, surf board visible in foreground.
- Multi-shot narrative. A 3-shot sequence: a person opening a door, walking into a room, sitting down at a desk. Same person across shots.
- Mixed dialogue + action. A character running and shouting “wait, don’t leave” over their shoulder. Combined motion and speech.
The results: 4 examples of Sora 2 output
Test 1. Cameo with dialogue
The Cameo feature is genuinely uncanny. After a 20-second consent video, Sora 2 produced clips where my face moved correctly with the dialogue, my hair behaved appropriately for the framing, and my voice approximation was close enough that a coworker recognised it before I told them what it was. The steam from the espresso machine drifted realistically across the frame.
What I liked
- Cameo likeness was production-grade
- Lip sync with my voice match was 4/5
- Environment (steam, light) reacted naturally
What I did not
- Voice clone was thinner than ElevenLabs
- Eyes occasionally drifted on long takes
- Cameo consent flow adds 1-day delay first time
Test 2. Physics-heavy action
This is where Sora 2 pulls ahead of Veo 3. The wave broke against the pier with realistic spray patterns; the foam dissipated correctly; the board in the foreground responded to the swell underneath. None of the takes had the typical AI-video tells (objects passing through each other, foam appearing then vanishing in a single frame). Out of five generations, four were postable to client work.
What I liked
- Best-in-class water and foam simulation
- Foreground/background depth was correct
- Light and color graded coherently
What I did not
- Long clips (over 10s) showed minor drift
- Rocks looked plastic in 2 of 5 takes
- No control over wave timing inside the clip
Test 3. Multi-shot narrative
The Sora app’s storyboard mode handled this cleanly. Across three shots, my prompted character kept the same hair, the same wardrobe, and a face that was recognisably the same person. This is where Sora 2 beats Veo 3 the most clearly: multi-shot identity locking. The desk in the third shot was a different desk than I imagined from the prompt, but the human was consistent.
Test 4. Mixed dialogue + action
A running character shouting over their shoulder is the unforgiving test for native audio. Veo 3 sometimes drops the audio out of sync when the head turns. Sora 2 kept the line lined up with the mouth even through the over-the-shoulder turn. The voice quality was thinner than a dedicated VO, but the timing was right. This is the test that historically forced you into a post-production lip-sync pass; Sora 2 makes it a one-shot generation.
The feature everyone is talking about: Cameos
The Cameos feature is the reason the Sora app went viral on launch day, and it deserves the attention. The consent flow is well-designed: you upload a short video, confirm verbally that you authorise its use, and OpenAI processes it before you can use it as a cameo. You can revoke at any time, and the cameo is yours alone unless you explicitly share it.
The practical use cases are wider than the meme footage suggests. Brand spokespeople can shoot once and appear in dozens of campaign variations. Course creators can record one segment and re-cast themselves into multiple language versions. Founders can generate explainer clips with their own face without a studio setup.
The caveats are the obvious ones: Cameos lives inside the Sora app and the OpenAI ecosystem. You cannot embed it in a third-party product or stitch it into an automated workflow without significant engineering. For pipeline work, you still want a flexible API-first platform on the side.
The annoying parts
$200/mo for the best tier. ChatGPT Pro is the only path to the longest clips, the highest-quality tier, and (in some regions) no visible watermark. For occasional users, that is steep.
Tiered API. Sora 2 API access rolls out based on organisation usage history, ID verification, and audio-enablement steps. Three weeks after my org was approved, audio was still gated. Plan procurement around this if you are integrating into a product.
App-first. The best Sora 2 experience is the consumer app. If you want Sora-grade output inside a production pipeline, you spend more time on glue code than on creative work.
Cameo collisions. When two cameos appear in the same scene, identity drift between them is noticeable. Useful for sketch comedy, problematic for branded multi-presenter videos.
API pricing math. The API bills Sora 2 at around $0.30 per second for the HD tier and around $0.50 per second for Sora 2 Pro. A 60-second client edit runs $18 to $30 per attempt. Five attempts is $90 to $150 before you have a usable clip.
Is it worth the price?
For a creator using Sora 2 daily inside the app, ChatGPT Plus at $20/mo is genuinely a steal. The Pro tier at $200/mo is a small-business expense and the right call if Sora 2 is your primary creative tool.
For an agency or product team, the math is harder. At $0.30 to $0.50 per second, per-second API pricing and tiered rollouts make capacity planning unpredictable, and the consumer-app focus means features ship to the app before the API. A flat-rate aggregator is often the cleaner integration path.
How Vuela.ai fits into a Sora 2 workflow
Sora 2 is the strongest consumer-facing AI video product on the market. It is not, by design, an integration platform. The jobs Vuela.ai handles sit outside the Sora app and outside the current API surface: cloning a viral video so you can ship a new version, translating finished video into other languages with real lip sync, and repurposing one render across a dozen aspect ratios.
Vuela.ai bundles Sora-grade video generation with that broader toolkit under a flat-rate subscription. If you came here because Sora 2 piqued your interest but you want a workspace that actually ships content, Vuela is what you stitch on top.
The verdict
Sora 2 is the most exciting AI video product to ship in 2025. As a model, it competes with Veo 3 on quality and pulls ahead on physics, length, and personal video (Cameos). As a product, it lives inside an OpenAI-built app that prioritises consumer use over pipeline integration.
If you want to play, the Sora app is the easiest fun you can have with an AI in 2026. If you want to ship content at volume, treat Sora 2 as one model in a larger workflow, alongside Veo 3 for audio-heavy clips, Kling 3 for identity-locked sequences, and a platform like Vuela.ai for the cloner and translator work that actually moves the needle.