AI video generators compared: Sora vs Veo vs Runway

The AI video race has entered its consolidation phase. As of September 2026, the three names everyone compared — OpenAI's Sora, Google DeepMind's Veo, and Runway's Gen-4 line — are no longer on equal footing. One of them is about to be switched off entirely.

This comparison runs all three through the same set of demanding test scenes — the kind of shots that separate real physics from plausible-looking fakery: a crowded café, a slow-motion splash, a drone flight, legible text in frame, and consistent characters across cuts. Where published hands-on tests and leaderboards exist, we lean on them. Where the models' own documentation draws the line, we say so. The goal is simple: which tool earns a place in your workflow in late 2026?

The big picture: a very different landscape than last year#

The most important column is status. OpenAI has confirmed in its help center that the Sora web and app experiences were discontinued on April 26, 2026, and the Sora API shuts down on September 24, 2026 — four days from now. Whatever Sora's strengths, it is not a tool to build on today. Treat everything below about Sora as a retrospective; the real race is Veo vs. Runway.

The test scenes: realism and motion#

To keep this honest, here's what each scene is designed to punish, and how the three models fare based on published evaluations, independent reviews, and leaderboards like Artificial Analysis:

Scene 1 — The crowded café. A wide shot of a busy coffee shop, waiters threading between tables, background extras in motion. This is the classic stress test: hands, faces in profile, overlapping bodies. Sora 2 was consistently strong here — its diffusion-transformer architecture handled multi-person scenes with fewer of the limb artifacts that plagued early models. Veo 3.1 is rated just as well for general cinematic realism, though reviewers note occasional artifacts in complex human motion. Runway's Gen-4.5 earns top marks for visual fidelity on benchmark leaderboards (Elo scores in the 1,240s on public leaderboards) and its character-consistency features keep extras from morphing between frames.

Scene 2 — The slow-motion splash. A coffee cup tipping over, liquid arcing through the air at high frame rate. Water and fabric are where Google DeepMind's physics background shows: Veo 3.1's fluid simulation is repeatedly cited as its standout strength — water drapes, splashes, and refracts light like something filmed with a Phantom camera. Sora 2's physics were also strong and, in OpenAI's own demos, quite cinematic. Runway holds its own on splash dynamics but is a notch behind Veo on the water itself — the difference between "film reference" and "very good mockup."

Scene 3 — The drone flight. A single continuous drone shot over a coastal road, camera swooping smoothly with the landscape. All three handle camera language well at this point — prompting a "slow push-in" or "aerial orbit" reliably produces coherent moves. Runway has an edge here, not in raw generation quality but in control: its camera-motion presets and motion brush let you direct specific elements of a shot instead of re-rolling prompts until the framing lands. Veo 3.1 counters with video extension — you can grow an 8-second clip shot-by-shot with reasonable continuity.

Scene 4 — Legible text in frame. A neon shop sign, a book cover, a license plate. This remains the cruelest test for video models. Veo 3.1 made real progress here — its 3.1 release specifically improved text rendering, and short phrases on signs now come out readable more often than not. Sora 2 and Runway remain inconsistent: expect mangled letters on longer strings, and plan on compositing real text in post if it has to be perfect.

Scene 5 — Same character, five shots. A recurring character across a sequence of generated clips. Runway wins this outright with its reference-image locking: feed it a character reference and it holds identity across shots better than either rival, which is why it's the default choice for narrative shorts and pre-vis. Veo 3.1 supports up to three reference images and first/last-frame specification, which is a genuine tool for keyframed sequences. Sora 2 had character support through its storyboard features, but with the app gone, it's a moot point.

Editability: where Runway laps the field#

If your work doesn't end at generation, this section decides your tool:

• Runway is a studio, not just a model. Motion brush lets you paint which regions of a frame move and how. Act-Two transfers a real human performance onto any character. Aleph, its AI editing layer, trims, retimes, and makes targeted edits (swap that background, change that color) without regenerating the clip. It also lets you call third-party models — Kling, Veo 3.1, Seedance — from inside one interface. Nothing else on this list is close for post-generation control.
• Veo 3.1 offers video extension (continue a generated clip), first-frame/last-frame interpolation, and image-to-video from reference images — solid production aids, but it is fundamentally a "generate the clip" tool. For a polish pass you export to your editor.
• Sora 2 had remix and storyboard-style features in its app, and the underlying model remains strong, but the app is dead and the API follows within days. Do not build a workflow on it.

Pricing and access, September 2026#

Pricing shifts constantly, so treat these as the verified-as-of-September-2026 snapshot and check the vendors' pages before committing:

• Veo 3.1 is the cheapest way to try anything serious. A genuinely free tier exists through Google AI Studio (rate-limited, up to 720p with audio), which makes Veo the only one of the three you can properly evaluate without paying. Paid access runs through Google AI Ultra (~$250/month, including 4K and priority queues), pay-as-you-go AI Studio billing (roughly $2 per clip), or the developer API, where Standard mode runs around $0.20/second and audio-enabled 4K goes higher.
• Runway runs on credits. The free tier gives 125 one-time credits (watermarked). Standard is about $12/month (annual) with 625 monthly credits, Pro about $28/month with 2,250, and Unlimited about $76/month annual. As a rough conversion, 625 credits buys roughly 25 seconds of flagship Gen-4.5 generation — so a Standard plan covers a handful of finished shots per month, not a volume pipeline. Use Turbo for drafts (much cheaper per second), Standard for finals.
• Sora is legacy. When it was available, access started at $20/month via ChatGPT Plus; the API priced at roughly $0.10–$0.50 per second. The question of whether it was worth it is now academic.

The verdict#

• Best overall for realism with audio: Veo 3.1. Native audio, strong physics, a genuinely free tier, and 9:16 native vertical for short-form. The 8-second clip cap is its main limitation — plan on stitching.
• Best for creators who edit: Runway. No other tool lets you art-direct, perform, and refine inside one product. The credit math punishes heavy volume, so it's best for deliberate, polished work rather than batch generation.
• Sora: an honorable mention, and that's it. OpenAI's model proved what a diffusion-transformer could do for video, and Sora 2's synchronized audio was genuinely ahead of its time. But with the API shutting down on September 24, 2026, recommending it for new work would be irresponsible. If you have Sora content to save, export it now — OpenAI will permanently delete associated data after the sunset window.

Takeaway#

The comparison that mattered in 2025 — which model generates the prettiest clip — has given way to a workflow question. Veo 3.1 is the default for fast, cheap, audio-included video; Runway Gen-4.5 is the default when you need to direct and refine; Sora is the default example of why you shouldn't build a business on someone else's research demo. If you're starting today: try Veo free in Google AI Studio this afternoon, and reach for Runway the moment you need control over what you generated.