The text-to-video landscape in 2026: Veo, Runway, Kling, Luma and life after Sora

0:00 The text-to-video landscape, 2026

In March twenty twenty-six, OpenAI announced it was shutting down Sora, one of the most famous video generators in the world. The app closed in April, and the models were removed from the API in September. If your workflow depended on it, you had a problem. In this lecture you'll get a clear snapshot of the text-to-video landscape as of September twenty twenty-six. More importantly, you'll learn a five-test evaluation you can rerun every quarter, and you'll generate a clip through an API yourself.

0:37 Why evaluate, not memorize

Why focus on evaluation rather than a leaderboard? Because this is the fastest-moving part of AI media. New model generations ship every few months, native audio went from rare to common, and a product can disappear. If you memorize today's rankings, you'll be wrong by the next quarter. If you know how to test tools fairly against your own needs, you'll make good decisions no matter what launches next. That's a skill clients and employers will pay for.

1:11 Snapshot: September 2026

Here's the snapshot. Google's Veo three point one family, including Fast and Lite variants, is available in the Gemini app and Flow for creators, and in the Gemini API and Vertex AI, which Google now calls the Gemini Enterprise Agent Platform. It generates native audio and supports reference images and first and last frame control. Runway Gen four point five is available in Runway's app and API. Kling three point zero from Kuaishou emphasizes multi-shot sequences. And Luma's Ray three family offers keyframe control and video modification. Verify each on the vendor's site before a project.

1:53 Lessons from Sora

And the Sora story teaches two lessons. First, any generator can disappear or change its terms, so keep your prompts, reference images and edit projects portable. Think of your tools like rented equipment in a film studio. You can rent the best camera in town, but your script, storyboard and footage should always live in your own cupboard. Second, export and archive everything you need. When a product shuts down, assets left inside it can be deleted. Portable inputs and archived outputs protect you.

2:30 The five tests

Now the five-test evaluation. Run the same five prompts on every candidate, same aspect ratio, same duration. Test one, concept b-roll. Test two, image-to-video from your actual product photo. Test three, consistency of a character or object across two shots. Test four, motion and physics, like pouring liquid. And test five, a scene from your market, like a Karachi street at dusk or a Riyadh office, checking for errors and stereotypes. Then score blind, meaning the reviewers don't know which tool made which clip.

3:07 Scoring and practical facts

Score each clip from one to five on prompt adherence, artifacts, motion realism, consistency, and audio if you use it. Then record four practical facts: time to first result, license terms, how the tool marks its output, and cost per usable second. That last one surprises people. If a cheaper tool needs six generations to get one usable clip, and a pricier one needs two, the pricier one may be cheaper in practice. Track the usable seconds, not the sticker price.

3:42 Example 1: a cooking creator

First example, a simple one. A creator wants vertical b-roll for cooking Shorts. She runs tests one and four on two tools she already has access to, with the same chai-pouring prompt. Tool A looks gorgeous but the liquid pours upward in two out of three takes. Tool B is slightly softer but pours correctly every time. She picks tool B for motion shots and keeps tool A for static mood shots. Two tests, twenty minutes, and a decision grounded in her own content.

4:19 Example 2: a Dubai agency

Second example, a business case with illustrative results. A Dubai agency produces b-roll for twelve clients. Every quarter, two editors run the full five tests on three candidate tools and score blind. In the latest round, one tool won concept b-roll, another won product image-to-video, and none handled culturally specific Gulf office scenes without errors, so those are still filmed. The agency keeps two subscriptions and writes a one-page tool policy that account managers share with clients. Clients like knowing the choice was tested, not guessed.

4:56 Watch me do it, part 1

Watch me do it through an API. I use Google's official Python SDK. I create a client with my key from an environment variable. I call generate videos with the Veo model id from the docs, and a director-style prompt that also describes the sound: quiet room tone, a gentle pour. I set the aspect ratio to nine by sixteen. Video generation is a long-running operation, so the call returns immediately with an operation, and I poll it every fifteen seconds until it's done.

5:33 Watch me do it, part 2

When it's done, I check for an error first, because a safety filter or quota limit can stop a job. If it's clean, I download the generated file and save it as chai opener dot mp4. Then I do the part people skip. I open my production log and write the model id, the full prompt, the date, and a note that Google marks Veo output with an invisible SynthID watermark. Preview model ids change, so that log is how I'll reproduce or explain this clip in six months.

6:12 Common mistakes

Common mistakes. Choosing a tool from its showcase reel instead of your own tests. Showcases are cherry-picked by definition. Ignoring cost per usable second. Building your whole workflow on one product with no export plan, which is exactly what hurt Sora users. And stripping watermarks or content credentials to avoid platform labels. That undermines trust, can break platform rules, and runs against the direction of regulation like the EU AI Act's transparency rules, which apply from August twenty twenty-six.

6:46 Recap

Recap. The landscape as of September twenty twenty-six includes Veo three point one, Runway Gen four point five, Kling three, and the Luma Ray three family, and Sora has been discontinued. Don't memorize rankings. Run the five tests, score blind, and compare cost per usable second. Keep your inputs portable and archive your outputs. Log model ids, prompts and dates. Looking ahead, and this is outlook rather than prediction, expect longer multi-shot clips, native audio as standard, and more provenance marking.

7:21 Try this now

Try this now. Pick two video tools you can access today. Run tests one, two and five from the lesson with identical settings. Ask a colleague to score the clips blind, and calculate cost per usable second. Write the result as a three-line tool policy: which tool for which job, and what you'll still film for real. Put a date on it and a reminder to rerun the tests next quarter. Then, if you're comfortable with a little code, run the Veo script from the lesson with your own prompt, and log the model id, the prompt and the date. You'll have both a decision and a working API pipeline by the end of the day.