Hands-On With Grok Imagine Video: 13 Clips, Nine Tests and a $3 Bill for a Tech-News B-Roll Desk

Three of the 13 clips from our test: a 720p server-room shot, a two-reference product shot and a presenter speaking a scripted line.

Tech news runs on B-roll. A story about chip prices needs a server aisle, a phone launch needs a product on a desk, and a weekly video bulletin needs a presenter who can say a line on cue. Text-to-video models promise custom footage on demand. The useful questions are practical ones: how long a clip takes, what it costs per second, whether the model does what the prompt says, and where it breaks.

We tested xAI's Grok Imagine Video model, grok-imagine-video, through its API with those questions in mind. The scenario is fictional: a small daily video bulletin we called Circuit Brief, which covers chips, phones and cloud infrastructure. We gave the model the jobs such a desk would give it: short B-roll, a multi-shot sequence, physical action, a product animated from a still photo, a presenter line with audio, a re-grade of an existing clip and an on-screen headline. We generated 13 clips, sent seven invalid requests on purpose, and logged the render time and billed cost of every job. The total bill for video was $3.02. Two input stills cost another $0.006.

What the model offers

The base model has three modes behind one endpoint. Send a prompt and you get text to video. Add image_urls and the model animates one or more reference images. Send a video.url with an instruction and it edits an existing clip. Clips run from 1 to 15 seconds in whole seconds, at 480p (the default) or 720p, in seven aspect ratios from 16:9 to 2:3. Prompts can be up to 8,000 characters. A separate 1.5 version adds 1080p but does not support editing; this test covers the base model only.

Pricing is per second of output, which matters more than usual here because duration is a free choice. The rates we were billed were $0.036 per second at 480p and $0.0504 per second at 720p, plus $0.0014 for each reference image. A five-second 480p draft therefore costs $0.18, and a fifteen-second 720p clip would cost $0.756. Every number in this article was checked against the cost field of the finished task.

How we called it

Generation is asynchronous. A POST returns a task ID, and you poll a task endpoint until the status is completed or failed. The video URL in the result expires after about a day, so download it straight away. The helper below is close to what we used. The host and key come from environment variables, and the task ID is written to a log file before polling starts. That last detail is not decoration: if your script crashes while polling, you can pick the task up again instead of submitting it twice and paying twice.

import os, time, requests

BASE = os.environ["VIDEO_API_BASE"]          # e.g. https://your-provider.example/v1

HEADERS = {"Authorization": f"Bearer {os.environ['VIDEO_API_KEY']}"}

def grok_video(name, prompt, seconds=5, resolution="480p", ratio="16:9",

               images=None, edit_url=None):

    if edit_url:   # editing takes only model, prompt and video

        payload = {"model": "grok-imagine-video", "prompt": prompt,

                   "video": {"url": edit_url}}

    else:

        payload = {"model": "grok-imagine-video", "prompt": prompt,

                   "duration": seconds, "resolution": resolution,

                   "aspect_ratio": ratio}

        if images:

            payload["image_urls"] = images     # public HTTPS URLs

    r = requests.post(f"{BASE}/videos/generations", headers=HEADERS,

                      json=payload, timeout=60)

    r.raise_for_status()                       # invalid parameters fail here

    task_id = r.json()["data"][0]["task_id"]

    with open("tasks.log", "a") as log:        # log before polling

        log.write(f"{name}\t{task_id}\n")

    while True:

        time.sleep(8)

        task = requests.get(f"{BASE}/tasks/{task_id}", headers=HEADERS,

                            timeout=30).json()["data"]

        if task["status"] in ("completed", "failed"):

            break

    if task["status"] == "failed":

        raise RuntimeError(task["error"]["message"])

    url = task["result"]["videos"][0]["url"][0]

    with open(f"{name}.mp4", "wb") as f:

        f.write(requests.get(url, timeout=300).content)

    return task["cost"], task["actual_time"]

We submitted the ten text-to-video jobs at once and the three image and edit jobs a few minutes later. All 13 completed on the first attempt. Render times below are the server's actual_time for each task, which excludes the time spent downloading.

Test 1: duration against time and cost

The first question for a news desk is what a longer clip costs in money and waiting time. We sent one prompt at three lengths, all at 480p and 16:9:

Slow dolly along a server-room aisle: rows of black racks with blinking green and amber status LEDs, a technician in a navy hoodie walks away from the camera at the far end and turns left, cold blue overhead light, faint haze, steady camera.

The same prompt at 1, 5 and 10 seconds. The 1-second clip compresses the whole action; the 10-second clip spends its extra time at a rack.

The one-second clip rendered in 24 seconds and cost $0.036. The five-second clip took 54 seconds and cost $0.18. The ten-second clip also took 54 seconds and cost $0.36. Cost scaled exactly with length, as the price list says. Render time did not: beyond a few seconds of output, doubling the length added no waiting time in our run.

The model fitted the action to the time it had. In one second the technician walked and began to turn, all in a hurry. In five seconds he walked, then turned to the right toward a rack, and in the last second his body blended into a rack door, a small but visible artefact. In ten seconds he walked, stopped and reached for a rack, and the motion was calm throughout. None of the three turned left as the prompt asked. For B-roll this hardly matters, but it is an early sign of a pattern we saw again: the model gets the scene and the mood right and is looser with specific directions.

Test 2: 480p against 720p, and a 12-second multi-shot prompt

We rendered the same server-room prompt at 720p. It took 76 seconds against 54 and cost $0.252 against $0.18, so 720p is 40 percent more expensive per second. The output was 1280×720 against 848×480.

480p and 720p versions of the same prompt, with enlarged crops of a comparable area of the racks.

The 720p version is not just sharper. The individual LEDs are separate points instead of soft streaks, the hoodie has texture, and the composition is more symmetrical. It was also the only one of our four server-room clips in which the technician actually turned left and walked out of the aisle. That is one sample, so we would not read much into it. For a social feed viewed on a phone, 480p is fine. For a 1080p bulletin timeline, where the clip will be upscaled, start from 720p.

The second half of this test used our one long clip. We wrote a shot list into a single prompt to see whether the model can deliver a short sequence, not just a single shot:

Three shots with hard cuts. Shot 1 (0-4 s): close-up of a hand placing a thin silver laptop onto a walnut desk, warm morning light. Shot 2 (4-8 s): over-the-shoulder view as the laptop lid opens and the screen lights up with a plain blue gradient. Shot 3 (8-12 s): wide shot of a small, bright home office; the person leans back in the chair and sips from a white mug. Quiet room tone, soft keyboard clicks, no music, no on-screen text.

The 12-second sequence. Shots one and two merged into one continuous move, the last change was a dissolve, and the laptop gained a logo nobody asked for.

This was the slowest job of the test at 169 seconds, and it cost $0.605. The content was mostly right: a hand on a silver laptop, the lid opening to a blue screen, then a wide shot of a home office in which the person sips from a white mug. The structure was not. There were no hard cuts. The first two "shots" became one continuous camera move around the 3-second mark, and the change to the wide shot was a one-second cross-dissolve at about 6 seconds, so the clip split into two halves of six seconds instead of three shots of four. The room tone was quiet and there was no music, as requested.

The bigger problem for a news site was on the lid: a fruit-shaped logo very close to a well-known trademark. The prompt said "thin silver laptop" and nothing about a brand. A news desk cannot publish B-roll with a recognisable manufacturer's logo next to a story that is not about that manufacturer. We saw the same mark again in Test 3. If you need multi-shot sequences, generate each shot as its own clip and cut them in an editor, and add "unbranded, no logos" to every prompt that shows a device.

Test 3: liquids and air

Physical plausibility is where video models tend to give themselves away. We ran two five-second 480p clips. The first poured water from a glass jug into a tall glass beside a closed laptop, asking for the glass to fill to the brim and spill a little. The second lifted a small drone off a cluttered desk and asked for its downwash to blow three sheets of paper off the edge while a potted plant's leaves shook.

Top: the pour fills the glass, but the puddle appears without a visible overflow. Bottom: the drone lifts off and one sheet blows away, then a sheet appears that was not there before.

The pour looked right at a glance. The stream bent naturally, bubbles rose, the jug's reflection moved with it, and the water level climbed steadily. But the glass never quite reached the brim, finishing at about 85 percent, and a puddle formed at its base anyway, with no water running down the side. The laptop next to it carried the same fruit-shaped logo we saw in Test 2. Both clips took 54 to 56 seconds and cost $0.18.

The drone clip was better on motion and worse on counting. The drone lifted off cleanly, with a slight wobble, and climbed out of the top of the frame. The desk had two loose sheets, not three. One of them flipped up and slid toward the edge in a believable way. Near the end, though, a sheet appeared on the desk mat where there had been none, and the plant's leaves stayed still the whole time. Under narration at broadcast speed these errors are easy to miss; in a hero clip they need checking frame by frame.

Test 4: a seven-point instruction prompt

To measure prompt adherence rather than describe it, we wrote a prompt with seven checkable instructions:

On a white table, from left to right: three blue cubes in a row and one red sphere on the far right. A robotic arm enters from the top of the frame, picks up the red sphere and places it to the left of the blue cubes. Then a small round lamp at the back of the table switches from red to green. Static camera, 45-degree top-down view, clean studio light.

The arm lifts the sphere and sets it down between the first and second cubes. The lamp changes colour before the sphere is released.

We scored the five-second 480p result, which took 54 seconds and cost $0.18, against the seven instructions:

Instruction	Result
Three blue cubes	Yes, exactly three
Red sphere on the far right at the start	Yes
Arm enters from the top of the frame	Yes
Arm picks up the sphere	Yes, with a clean grip
Sphere placed to the left of the cubes	No: placed between the first and second cube
Lamp switches red to green, after the move	Partly: it switched, but while the sphere was still in the gripper, and it hangs above the table instead of standing at the back
Static camera	Yes

That is five clear passes, one partial and one miss. Counting and object identity were solid. Placement and the order of events were the weak points. If your clip has to show a specific sequence, say a product being plugged in and then lighting up, describe one action per clip and join them in the edit, or animate a still frame that already shows the end state, as in the next test.

Test 5: animating a product still

A news desk often has one press photo and needs motion. We made a stand-in: a 1K still of a fictional sage-green phone we called the Halcyon One, with a two-lens camera bump and the brand name engraved small on the back. It cost $0.003 from an image model. We then passed its URL in image_urls of a request to the Grok Imagine Video API, with this prompt:

A hand enters from the right, picks up the phone, turns it over to show its dark glass screen, then places it back face down in the same spot. The camera stays still. Soft daylight, quiet room.

Top: one reference image, animated. The phone ends where it started, with the engraved name still legible. Bottom: phone plus style reference.

This was the best result of the test. The clip rendered in 33 seconds, the fastest of any five-second job, and cost $0.1814: five seconds at the 480p rate plus $0.0014 for the image. The first frame matched the still. A hand picked the phone up, turned it to show a dark screen, and put it back. In the last frame the phone sits in almost exactly its original position, and the small engraved wordmark is still readable. The lamp, plant and book in the background stayed put throughout.

Two details are worth knowing. First, the output was 736×400 pixels rather than the 848×480 of our text-to-video clips, so the image route gave a slightly smaller frame at the same price. Second, the phone's screen side was invented, because the still only showed the back. The model drew a plain black screen with thin bezels, which is reasonable, but a real review unit would need a second reference.

Test 6: a subject image and a style image

The API keeps the order of reference images, so you can assign them roles in the prompt. We sent the phone still first and a neon-noir rainy street, also generated for $0.003, second:

Use image 1 for the phone and image 2 for the look. The same sage-green phone lies face down on a wet window ledge at night; magenta and cyan neon reflections slide across its back as rain runs down the glass. Slow push-in toward the camera bump.

The clip took 44 seconds and cost $0.1829, which is the 480p rate plus two images. The look transferred completely: magenta and cyan neon, wet reflections, rain and haze. The push-in toward the camera bump worked. The phone kept its shape and two-lens bump, and the brand name was faintly visible at the end. Two things drifted: the sage green paled to mint under the neon, and there was no window ledge. It put the phone on a wet surface in the middle of the street from the style image. In other words, image 2 was used for its scene as well as its look. If you want only a colour grade, describe the scene in words and use a style reference that has no strong setting of its own, or apply the grade with the edit mode in Test 8.

Test 7: a presenter line with sound

Every clip we generated came back with an AAC audio track, so we tested whether the model can deliver a scripted line. The prompt described a presenter in a charcoal blazer at a studio desk and gave her a line in quotes: "Good morning, this is Circuit Brief. Chip prices fell twelve percent this week." We asked for six seconds at 480p, which took 54 seconds and cost $0.216.

The presenter line. A speech-to-text pass returned the script word for word.

We ran the audio through a speech-to-text model. It returned "Good morning, this is Circuit Brief. Chip prices fell 12% this week." That is every word, in order, including the invented programme name. Silence detection put the first sentence at 0 to 1.8 seconds, a natural pause, and the second sentence ending at 4.8 seconds, leaving 1.2 seconds of room tone at the end. The mouth movements matched the speech at normal playback speed, and the presenter added small hand gestures on the second sentence without being asked.

The clips without dialogue had ambient tracks: hum for the server room, a soft trickle for the pour, keyboard clatter for the typing shot. The speech-to-text model heard no real words in any of them. For short presenter lines the built-in voice is good enough to rough out a script before recording a real host.

Test 8: editing an existing clip

The edit mode takes a public video URL and an instruction. It does not accept a duration, resolution or aspect ratio; the output follows the source. We sent the five-second 480p server-room clip from Test 1 with this instruction:

Re-grade this clip as a warm sunset scene: orange light streaming in from the left, warmer shadows. Keep the camera move, the racks and the walking technician unchanged.

Source (top) and edit (bottom). The motion and layout held; the light moved to the far end of the aisle rather than the left.

The edit took 55 seconds and returned an 848×480 clip of the same length. The camera move, the rack layout and the technician's walk were unchanged frame for frame, which is what makes the edit mode useful: you can match a clip to a story's colour palette without losing a take you like. The light, however, came from the far end of the aisle as a glowing orange backlight, not from the left, and the LEDs kept their green colour. The cost was $0.288, which is more than the $0.18 we paid to generate the source clip. So editing is not a cheap way to make variations. It is a way to keep a specific take. If any take will do, a new generation with a warm prompt costs less.

Test 9: headline text, small text, logos and invalid requests

We finished with the jobs most likely to embarrass a news site. A five-second vertical clip asked for two hands typing on a keyboard with white keycaps and a red lower-third banner that "reads exactly" CHIP PRICES FALL 12%.

The banner was exact in every frame. The keycap letters were invented, and the laptop in Test 3 carried an unrequested logo.

The banner was perfect. Every character, including the percent sign, was correct in every frame, the bar stayed fixed while the hands moved, and the vertical framing was right. The keycaps were another story: rows such as "A S", "F K D" and "I K" show letters in the wrong places, and some keys repeat. Hands were convincing, with five fingers each and natural typing motion. This clip took 56 seconds and cost $0.18.

Large text that you quote is reliable. Small text that the model invents, such as keycaps, labels or screen content, is not, and neither are product details that resemble real brands. Put headlines in quotes, keep incidental text out of focus or out of frame, and check every device for logos.

We also sent seven requests that we expected to fail. Six were rejected immediately with an HTTP 400 and a readable message, and none created a task: a duration of 16 seconds, a duration of 0, 1080p on the base model, a 21:9 aspect ratio, an empty image_urls array and an empty prompt. The seventh, a reference image URL that does not exist, was accepted and failed 15 seconds later with a generic message suggesting a retry. It was not billed. The lesson for a pipeline is to check that your image URLs return an image before you submit them, because a failed fetch looks like a temporary problem and a naive retry loop will keep trying.

Speed and cost across the test

Render time and billed cost for every clip in the test.

The median render time was 54 seconds. Eight of the 13 jobs finished between 54 and 56 seconds, whatever their content. Image-to-video jobs were faster (33 and 44 seconds), the 1-second clip took 24 seconds, the 720p clip 76 seconds and the 12-second 720p sequence 169 seconds. Ten jobs submitted at once finished within seconds of each other, so parallel jobs did not queue at this volume.

The full test cost $3.02 for 13 clips, about $0.23 per clip. A daily budget for Circuit Brief of six 5-second 480p social clips (1.08)andtwo8-second720pbulletinclips(0.81) comes to about $2 a day, or roughly $45 for a 22-day month. The larger cost is the editor's time spent checking clips.

The walkthrough video

Narrated walkthrough of the nine tests with the generated clips (2 minutes 38 seconds).

Checklist for a news desk

1. Preview with 1-second clips. At $0.036 and about 24 seconds, they show the look of a prompt before you pay for length.
2. Pay for seconds, not for waiting. Cost scales with duration; render time barely does. Generate the length you will use.
3. Use 720p for anything that goes full screen. It costs 40 percent more per second and is clearly sharper.
4. Do not rely on hard cuts in one prompt. Our three-shot request came back as two halves joined by a dissolve. Generate shots separately.
5. Write "unbranded, no logos" for every device. Two of our laptop shots carried a fruit-shaped logo nobody asked for.
6. One action per clip. The model counts objects well but places them loosely and can change the order of events.
7. Animate a still when identity matters. Image to video kept our phone and its engraved name intact and was the fastest job of the test.
8. Expect style references to bring their scene along. Describe the setting in words if you want only the colours.
9. Quote headlines; avoid incidental text. The banner was exact, the keycaps were invented.
10. Validate image URLs before submitting. A missing image fails after the task starts, not at submission.

The bottom line

Grok Imagine Video handled most of a tech-news B-roll brief for about three dollars. It was strongest where a desk needs it most: atmospheric short clips with a consistent look, a product still brought to life without losing its details, an exact on-screen headline, and a presenter line delivered word for word with matching lip movement. It was weakest on structure and precision: shot lists, the order of actions, small invented text, physically consistent liquids and, most importantly for a publisher, device logos that appear unprompted. Those weaknesses are manageable with short single-action prompts, references for anything that must stay accurate and a careful look at every frame before publishing. If you want to repeat these tests with your own prompts, or compare Grok against other video models behind one key, you can find them on APIMart.