OpenAI pulled the Sora app earlier this year, and the tool everyone started pointing at was Grok Imagine.

It does video now, and it's fast and cheap enough that people are calling it the replacement, so I gave it three jobs and watched what it did.

The surprise: the real contest was Grok against itself.

When you generate a still in Grok Imagine and hit the video button, Grok offers its own animation first, its read on how the frame should move.

You can also ignore that and write your own motion prompt. I ran both on every test. Before that, some catchup on AI this week:

Partner with us

Output tokens run 3-5x the price of input, and a single image can cost more than 5 pages of text.

Most people never see where the money goes, so their bill climbs and tand most assume that's just the cost of using AI, which isn't true.

A few of the easy wins:

Six more inside, including the API-side plays: prompt caching, the /compact command in Claude Code, and a tool that strips coding-agent output by up to 90%.

Still prompt: Close-up portrait of a young woman in an empty cafe, looking just off camera, warm window light, shallow depth of field, natural skin texture.

Animate prompt: She blinks, glances toward the camera, and a small smile forms, her hair shifts slightly in a draft, camera holds steady.

Output:

Grok's instinct lowered her eyes and brought her hands toward her face, a small private moment that read like a thought passing through.

My prompt told her to turn to the camera and smile, and she did, with real warmth, though the direct eye contact tipped a little into eerie.

I couldn't pick a winner.

Grok's version carried more emotion, mine was the more striking shot, and both would survive as a hero frame.

Still prompt: Overhead shot of a glass of dark coffee, a thin stream of cream already falling into it, side lighting, steam rising

Animate prompt: The cream stream keeps pouring and blooms into a swirling marble pattern through the coffee, steam curling upward, camera holds steady

Grok's instinct let the cream act like cream. My prompt pushed it toward a controlled, near-symmetrical pattern, almost latte art.

Grok's automatic version was the more physically truthful of the two.

Mine bent the liquid into something liquid wouldn't quite do, but it looked more finished and deliberate, and that's the one I kept.

Still prompt: Matte black wireless earbuds case on a wet rooftop at night, neon reflections on the surface, brand text 'SOMETHING' embossed on the lid, product photography

Animate prompt: Slow cinematic push-in over the case, water droplets bead on the lid, neon reflections shift across the surface, camera moves steadily.

Grok beat me outright. Its own version flipped the lid open and lit the earbuds with a cyan ring a beat after they were exposed, which is what real earbuds do when they wake up.

My prompt asked for water droplets forming on the lid, and they showed up out of nowhere with no rain to explain them, physics wandering all over the place.

I'd asked for an effect the scene couldn't justify, and it faked it, badly.

Between Grok's instinct and my direction, the scoreboard came out level: one draw, and one win each. The auto-suggestion is a solid second opinion, and often a sharp one.

Grok's defaults are smart, sometimes smarter than the instruction.

So for your own work: let Grok take the first swing before you start steering. Treat its suggested motion as the draft and your prompt as the edit, not the other way round.

Save the heavy direction for when you want a particular polished look, and back off when you'd be forcing physics the image can't pay for.

And the big question, is this a real stand-in for Sora? For short clips, made fast and cheap, it holds up.

The pitch is the speed and the price, clips in seconds at around $4.20 a minute for 720p against Sora 2 Pro's $30, and the output mostly earns it.

It has tells but the floor is high, and on the face and the product shots the ceiling was higher than I expected.

If the last time you opened Grok Imagine it only made images, it's worth another look.

Until next time,
Vaibhav 🤝🏻