Grok Imagine Video test: useful visuals, shaky timing

In Quasa.io’s reported sample, Grok Imagine Video produced usable short-form atmosphere, but struggled more when shots had to follow exact timing, placement or transitions. A report published November 2, 2026, covered 13 clips made with the base grok-imagine-video model. Quasa.io put the run’s video costs at $3.02; xAI’s published direct API rates are a separate price basis.

What the test covered

Quasa.io tested the base model across nine task categories for a fictional technology-news workflow, generating 13 clips and submitting seven deliberately invalid requests. The report included text-to-video, image animation, editing, presenter speech, on-screen text and prompts with multiple actions.

The clearest takeaway is a split between visual mood and precise direction. Some clips delivered useful-looking footage or a legible main headline. In other samples, requested actions happened out of order, shots ran together, and small details changed. Those are observations from this set of clips, not a guarantee of how every prompt will turn out.

Atmosphere came more easily than exact choreography

Quasa.io’s server-room prompt asked for a left turn. In three 480p clips lasting one, five and 10 seconds, the requested turn did not happen. Those renders took 24, 54 and 54 seconds. In a separate resolution comparison, the 720p clip looked sharper and was the only one of four server-room clips to turn left; it took 76 seconds, compared with 54 seconds for the 480p sample.

More elaborate instructions proved harder to keep intact. A 12-second prompt specifying three four-second shots took 169 seconds to render. The first two shots merged into one move, and the last transition dissolved instead of cutting. The clip also included an unrequested fruit-shaped logo on a laptop.

In a five-second scene with seven instructions, Quasa.io counted five clear passes, one partial pass and one miss. The model identified and counted the objects correctly, but put a red sphere between two cubes rather than to their left. It also changed a lamp before the requested movement was finished.

Other clips showed similar trouble with exact physical details: a glass ended about 85% full despite a puddle appearing, and a drone scene changed the number of sheets and left plant leaves still when the prompt asked them to move. For a scene where the sequence or object placement is the point, these misses matter more than a polished-looking frame.

Reference images, speech and text

xAI documents text-to-video, image-to-video, video editing and video extension for the base model. Its image-to-video documentation lists public image URLs, data URIs and Files API IDs as accepted image inputs.

Quasa.io’s fictional-phone example kept the visible back and engraved name recognizable when animating a still, but invented the unseen screen side. Adding a second image as a style reference transferred some of that image’s setting too, and shifted the phone’s sage-green color toward mint. One successful still animation shows that the mode can produce a usable result; it does not tell you how consistently it will preserve details across other images.

A six-second presenter clip reproduced its requested line in a speech-to-text check, and Quasa.io judged the lip movement synchronized at normal playback speed. In another sample, a large headline rendered correctly, while small keyboard legends were garbled or misplaced. The report also describes unrequested laptop marks in more than one clip. Keep essential text, product details and branding under review when they need to be exact.

What xAI lists for the base model

xAI lists 480p and 720p output and a general duration range of one to 15 seconds for grok-imagine-video. Reference-to-video has a separate limit: the xAI reference-to-video documentation lists up to seven reference images, 720p and clips up to 10 seconds for the base model. xAI’s video API reference describes an asynchronous workflow: submit a request, check its status using the request ID, then retrieve the result.

These are capabilities and limits listed for the base model. Quasa.io’s performance findings concern its own base-model sample; they do not describe version 1.5.

The reported test bill and xAI’s direct rates

Quasa.io reports $3.02 in video costs for its 13 clips, plus $0.006 for two input stills. Its reported unit rates differ from xAI’s direct API list rates. xAI’s pricing page, last updated October 10, 2026, gives these USD rates for the base grok-imagine-video model:

At xAI’s listed rates, a five-second 480p clip costs $0.25 and a 15-second 720p clip costs $1.05, before any input-image charge. By comparison, applying Quasa.io’s reported rates gives $0.18 and $0.756 for those same durations and resolutions. The figures are useful only with their rate basis attached: the $3.02 total is Quasa.io’s reported test bill, not xAI’s listed direct-rate calculation.

Quasa.io also proposed a daily workload of six five-second 480p clips and two eight-second 720p clips. Using its stated test rates, that works out to 6 × 5 × $0.036 = $1.08 and 2 × 8 × $0.0504 = $0.8064, or $1.8864 per day. Across 22 days, the total is $41.5008, about $41.50 before other charges.

For a workflow built around quick atmospheric footage, the sample contains promising outputs. When timing, transitions, object positions, small text or product details have to be exact, the same report gives plenty of reasons to inspect each generated clip before using it.