Twelve months ago, AI video generation was still a novelty act — clips lasted a few seconds, faces melted if a subject turned their head, and anyone with a trained eye could pick the fakes in about half a second. That’s no longer true. In 2026, the gap between “obviously AI” and “wait, is that real?” has narrowed dramatically, and three tools are doing most of the heavy lifting: OpenAI’s Sora, Google’s Veo, and Runway’s Gen line. We’ve spent the last few weeks putting all three through their paces on the kind of briefs a small Australian business, content creator, or indie filmmaker would actually throw at them, and the results are worth a proper comparison rather than another breathless “AI can do anything now” write-up.
We covered the text and chat side of this shift when we compared ChatGPT, Claude, and Gemini for Australian users, and the same pattern is playing out in video: the underlying models are converging on quality, but the tools built around them differ enormously in what they’re actually useful for day to day.
The state of play in 2026
Sora, now fully rolled into ChatGPT’s paid tiers, remains the most cinematic of the three. It handles camera movement, depth of field, and lighting continuity better than anything else we tested, and it’s the one most likely to produce a shot that wouldn’t look out of place in a low-budget feature film. Veo, Google’s entry, is tightly bundled into the Google ecosystem — Gemini, Vertex AI, and increasingly YouTube’s creator tools — and its strength is grounding: it tends to respect physical plausibility and text-in-scene rendering more consistently than the others, which matters more than people expect once you’re trying to use generated footage for anything beyond a mood reel. Runway, meanwhile, has doubled down on being a working editor’s tool rather than a one-shot generator. Its Gen-4 pipeline includes motion brushes, camera controls, and multi-shot consistency features aimed squarely at people who are actually cutting a sequence together, not just generating a single clip and hoping for the best.
None of these are toys anymore. But none of them are a replacement for a camera crew either, and the gap between the marketing demos and what you get on a Tuesday afternoon with a real brief is still significant.
Output quality and realism
On raw realism, Sora edges ahead for human subjects in natural settings — skin texture, hair movement, and ambient lighting all look convincing at first glance. Veo is close behind and arguably more reliable for outdoor and nature footage, where its grounding in real-world physics shows. Runway’s raw output is slightly behind both on photorealism but makes up for it with control: you can guide a shot with a reference image, a motion path, or a rough storyboard frame, which matters enormously when you need a specific result rather than a pleasant surprise.
This mirrors what we’ve seen on the stills side too. When we looked at AI-powered phone photography, the pattern was the same: the models that look most impressive in a single generated frame aren’t always the ones that hold up under actual production use, where consistency and control matter more than a single hero shot. Video makes this problem worse, because you’re not judging one frame — you’re judging every frame, for the length of the clip, and the seams show up the moment something moves in a way the model hasn’t seen enough examples of.
Length limits and cost
Clip length remains the single biggest practical constraint, and it’s easy to forget when you’re only looking at showreels. Sora’s standard generations top out at around 20 seconds on the consumer plans, with longer sequences requiring stitching multiple generations together — which reintroduces the consistency problems these tools are meant to solve. Veo sits in a similar band, with Google’s enterprise tier via Vertex AI allowing longer, higher-resolution renders at a meaningfully higher price point. Runway allows longer continuous generations on its higher tiers, but quality and coherence noticeably degrade the further out you push, especially past the 30-second mark.
On cost, none of these are cheap once you’re generating at any real volume. Sora access is bundled into ChatGPT’s higher subscription tiers, which makes the per-clip cost hard to isolate but easy to blow through if you’re iterating on a brief. Veo pricing through Google’s consumer apps is comparable, while the Vertex AI enterprise path is priced per second of generated footage and adds up fast for anything beyond short social clips. Runway uses a credits system, and a single minute of finished, edited footage — accounting for the generations you’ll throw away — routinely costs more in credits than most small businesses expect going in. Budget for iteration, not just output: the number of generations you discard before landing on a usable clip is usually higher than the marketing material implies.
What they’re actually good for
Where all three genuinely earn their keep is short-form marketing and social content — product hero shots, abstract background footage, stylised transitions, and B-roll that would otherwise require a stock footage licence or a location shoot. A local business wanting a 10-second scene-setting clip for an Instagram ad, or an agency needing a dozen variants of a concept to test, is exactly the use case these tools were built for, and they’re genuinely faster and cheaper than the alternative.
Indie filmmaking is a more interesting case. We’re seeing small Australian production teams use Sora or Runway for previsualisation — generating rough versions of a scene to plan blocking and camera angles before a real shoot — and occasionally for cutaway shots that would be prohibitively expensive to film for real, like a wide aerial or a period-specific street scene. It’s rarely the star of the finished piece, but it’s a genuinely useful pre-production and cost-saving tool. The same instinct that’s pushed developers toward AI coding assistants like Copilot, Cursor, and Claude Code to handle the repetitive parts of a build is now showing up in small production houses using AI video generation to handle the repetitive parts of a shoot, freeing up budget and time for the shots that actually need a human crew.
Where they still fall apart
Hands remain the classic tell, though they’ve improved a lot — fingers merging or bending at impossible angles is rarer than it was, but still turns up reliably once you generate more than a handful of clips with visible hands doing anything specific, like holding an object or typing. Physics is the deeper problem: liquids, cloth, hair, and anything involving weight and momentum still behave approximately rather than accurately. Water that doesn’t quite splash the way it should, fabric that moves like it’s underwater, objects that fall at the wrong speed — these are subtle enough that a casual viewer scrolling a feed won’t clock them, but obvious the moment you’re looking for them.
Consistency across shots is the constraint that matters most for anyone trying to make something longer than a single clip. Ask any of these tools for a character in one scene and then the same character in a different scene, and you’ll get someone who’s recognisably similar but not identical — wardrobe details drift, facial proportions shift slightly, and lighting logic resets between generations. Runway’s reference-image and character-consistency tools are the best attempt at solving this so far, but “best attempt” is doing some work in that sentence — none of the three have actually solved it.
The deepfake and misinformation problem
The realism gains that make these tools useful for marketing are the same gains that make them a genuine problem for misinformation. A convincing 10-second clip of a public figure saying something they never said is now trivially achievable, and it doesn’t need to be perfect to do damage — it just needs to be convincing enough to be shared before anyone fact-checks it. All three providers have leaned on watermarking and provenance standards like C2PA to label AI-generated content, but those markers are metadata, not something visible in the footage itself, and they’re stripped the moment a clip is re-uploaded, re-encoded, or screen-recorded — which is most of the time, in practice.
Spotting AI video without relying on invisible metadata is still possible if you know what to look for: watch the hands and background details rather than the main subject’s face, look for lighting that doesn’t quite match between foreground and background, and check whether reflections, shadows, and text in the scene are internally consistent. eSafety’s guidance on deepfakes and digitally altered material is a genuinely useful starting point for anyone — parents, small business owners, or just regular social media users — trying to understand the practical risk and what to do if they encounter or are targeted by synthetic content. It’s aimed at a general audience rather than technical readers, which is exactly the point.
Licensing and copyright for Australian creators
This is the messiest part of the whole picture, and it’s not close to settled. All three providers’ terms of service grant you usage rights over what you generate, but none of them can fully guarantee the training data underlying the model was licensed the way you’d want it to have been, and Australian copyright law hasn’t caught up with generative video the way some other jurisdictions are attempting to. If you’re generating footage for commercial use — an ad campaign, a paid client project, anything with real money behind it — read the specific commercial-use terms for whichever platform you’re on, because they differ meaningfully between Sora, Veo, and Runway, and they change often enough that last year’s understanding may already be out of date.
There’s also a reputational angle that’s easy to overlook: using AI-generated footage that closely resembles a real, identifiable person without consent creates legal and ethical exposure regardless of what a platform’s terms of service technically permit. The Australian Government’s AI Ethics Framework lays out the practical principles worth applying here — transparency, fairness, and accountability — and they’re a reasonable checklist to run through before using generated footage in anything client-facing or public. If in doubt, disclose that footage is AI-generated. It costs nothing and it heads off the kind of trust problem that’s much harder to fix after the fact.
Final thoughts
None of these three tools is a universal winner, and anyone telling you there’s a single “best” AI video generator in 2026 is oversimplifying. Sora is the strongest choice when cinematic realism is the priority and you’re working within short clip lengths. Veo earns its place when grounding, physical plausibility, and integration with the Google ecosystem matter more than raw cinematic polish. Runway remains the most practical option for anyone actually editing a sequence together rather than generating a single standalone clip, thanks to its control tools and consistency features.
What hasn’t changed is the need for a human in the loop. These tools are genuinely useful for marketing content, previsualisation, and filling gaps a real shoot can’t afford to cover — but the hands, the physics, and the shot-to-shot consistency problems mean they’re still assistants rather than replacements. And the same realism that makes them useful is exactly what makes the misinformation risk real, which means disclosure, provenance awareness, and a healthy scepticism toward anything that looks a little too smooth are going to matter more, not less, as these models keep improving.
