Grok Imagine Video 1.5 spent only thirteen days under its announced preview label. xAI opened grok-imagine-video-1.5-preview on June 3, 2026, then declared grok-imagine-video-1.5 generally available in the Imagine API on June 16.

The short interval can make the releases look like a simple rename. They were more than that: the GA announcement defined a clearer production model, described quality and audio improvements, added a Fast product experience, and attached the release to workflow features in Grok Imagine. Yet several core boundaries stayed the same. Understanding both sides prevents a migration from silently expanding assumptions.

What the preview promised

The preview was a focused image-to-video API. A developer supplied one still image and a text prompt, then selected a clip duration and resolution. xAI described preservation of source detail and lighting, along with generated camera moves, atmosphere, physics, and sound design.

The launch post showed the model ID grok-imagine-video-1.5-preview, a ten-second example, and output up to 720p. It suggested generating sequences by staging multiple source frames, animating them separately, and chaining the shots. That was a workflow recommendation, not a claim that the model accepted a storyboard full of references in one request.

Preview status meant teams should expect interface, behavior, capacity, or pricing to change. A sensible preview integration would isolate the provider adapter, store every request ID, and keep the model behind a feature flag.

What GA added or clarified

The GA announcement centered on four changes.

A stable production name

The model became grok-imagine-video-1.5. The current model page lists the former preview name and a dated grok-imagine-video-1.5-2026-05-30 name as aliases. Existing calls may continue to resolve, but the GA ID communicates intent and is the better default for new production code.

An alias is not a substitute for version governance. Log the name sent, the identity returned when available, and the date of every regression run. If reproducibility matters, evaluate whether the dated alias has the stability semantics the application needs rather than assuming from its spelling.

Stronger motion and physics claims

xAI says 1.5 improves movement across a clip, producing fewer warps and more believable weight and momentum. This expands the product’s intended usefulness from a brief animated effect toward more complete creative shots.

The claim should still be tested on domain footage. Hair, fingers, fluids, packaging text, reflections, fast rotation, and occlusion stress different parts of a model. GA is an availability state, not a guarantee that all physics are correct.

Improved native audio and speech

The GA post says effects, ambience, and dialogue are generated in the same pass and better aligned to action. It specifically describes clearer speech and improved synchronization.

That can reduce the work needed for a concept preview. It also adds an independent failure channel. A clip can be visually accepted while speech is wrong, or audibly convincing while the image drifts. Production review must inspect both tracks and their relationship.

A faster app experience

xAI rolled out Grok Imagine Video 1.5 Fast on the web and mobile apps and reports about 25 seconds for a six-second 720p clip, down from more than 40 seconds for its previous model. This number belongs to the Fast product variant described in the launch announcement. It should not be treated as a guaranteed latency for every GA API call.

The same announcement introduced Projects, parallel agents, and search for the Grok Imagine workspace. Those improve organization around generations; they are not parameters added to the video model itself.

Current resolution and pricing are broader

The preview announcement said up to 720p. The current GA model page lists output pricing for 480p, 720p, and 1080p: $0.08, $0.14, and $0.25 per second respectively, plus $0.01 for an image input at published rates.

That makes high-resolution cost easy to underestimate. Ten seconds at 1080p is $2.50 plus input for one request; multiple creative takes multiply that figure. Recalculate budgets and rate limits rather than carrying preview assumptions forward.

Use a two-stage policy when possible. Develop composition and motion at a lower tier, then promote a locked source and prompt to final resolution. Compare accepted output rather than assuming a higher pixel count fixes temporal or semantic errors.

What did not change

Grok Imagine Video 1.5 remains a model with image input and video output. Its current page explicitly says it does not support text-to-video. xAI’s reference-to-video guide says version 1.5 does not support that mode either.

The broader grok-imagine-video model can support text, image, and video workflows described elsewhere in xAI’s documentation, but those capabilities should not be copied into a 1.5 request. Model selection must happen before prompt design.

The API also remains asynchronous. A REST application submits a request, persists the returned ID, polls for completion, and stores the resulting video. Generated URLs can expire, moderation still applies, and outputs remain nondeterministic.

A migration checklist

Moving from preview to GA should be a controlled release:

  1. replace the preview name with grok-imagine-video-1.5 in a staged environment;
  2. verify the exact request schema used by the current SDK or REST API;
  3. regenerate a representative set of source images and prompts;
  4. compare identity, motion, physics, text, dialogue, and audio sync;
  5. recalculate cost for every supported resolution and duration;
  6. confirm timeouts, polling intervals, terminal states, and rate limits;
  7. download results into durable storage and verify checksums;
  8. update monitoring dashboards to split preview and GA results;
  9. roll out gradually with a rollback path.

Do not compare only the “best” output from each model state. Use the same number of attempts and record selection rate. Otherwise, an apparent quality gain may be more sampling.

For a high-volume migration, shadow traffic is safer than an immediate switch. Send a small authorized sample to both the old preview alias and the GA ID without publishing duplicate outputs. Compare status distribution, latency, media metadata, and reviewer scores, then stop dual generation once confidence is sufficient. Shadowing costs money, so set a fixed sample and expiry rather than leaving it enabled indefinitely.

Update operational alerts as well as creative tests. A new model can change queue time, error messages, output file size, or the time before a URL expires. Dashboards should show completion rate and download success separately; otherwise a generation may appear healthy even though downstream storage is failing.

Revisit the audio decision

Better native audio changes downstream work but does not eliminate it. Classify every accepted clip:

Transcribe dialogue and check claims, names, and numbers. Review loudness, clipping, channel layout, and timing. Record whether audio was model-generated or added later so editors and publishers understand provenance.

Keep a silent fallback rendition when the visual is independently useful. Social platforms and browsers apply different autoplay rules, and accessibility or localization may require a separate track. Treat the generated audio as one rendition attached to an approved picture master, not as an inseparable truth about the clip.

Revisit cleanup rather than assuming it disappears

Improved source fidelity may reduce warping, but GA can change the type of cleanup instead of removing it. A more coherent clip may still include unwanted text, accidental subtitles, unusable sound, or a frame that fails a brand rule.

Medux can be an independent finishing service called by an authorized Codex or Claude MCP client; it is not native to xAI. If a good visual contains removable unwanted subtitles, the Codex MCP subtitle-removal tutorial illustrates that separate task. If the native audio is rejected, the Claude MCP audio-track tutorial shows a replacement workflow.

Keep the xAI request and Medux task as separate records joined by the approved output checksum. A cleanup retry should not regenerate the Grok clip, and a Grok retry should not overwrite an approved cleanup artifact.

The preview-to-GA transition makes Grok Imagine Video 1.5 easier to adopt with a production contract, higher documented resolution, and clearer performance claims. It does not excuse weak state handling or broaden the model beyond one-image animation. The best migration recognizes the improvements, preserves the limits, and measures whether GA actually lowers cost per accepted, publishable clip.