Grok Imagine Video 1.5 API: verify reference images and 1080p before rollout
Quick answer
xAI announced on July 31, 2026 that grok-imagine-video-1.5 now supports text-to-video, image references, and native 1080p in the API. The release example passes reference_image_urls, allows up to seven visual references, and says voice references are available in the API on request.
Do not treat every headline capability as a stable contract for every account yet. xAI's older reference-to-video guide, last updated July 27, still says the 1.5 model does not support that mode, while the newer release page says it does. The general video guide also limits 1080p to image-to-video even though the release says 1080p works for text-to-video too. The safest rollout is therefore:
newer release contract
-> account-level capability canary
-> output and cost verification
-> bounded production rollout
This guide supplies the capability matrix, cost calculation, and eight tests needed for that gate.
Who this is for
This is for developers integrating Grok Imagine Video 1.5 into an app, content pipeline, product-demo generator, or automated creative workflow. It is especially useful if you are searching for reference-image support, 1080p pricing, text-to-video availability, or the difference between grok-imagine-video-1.5 and the broader grok-imagine-video workflow.
If an agent will generate assets automatically, keep the same bounded-job principles used in the Grok Build parallel-agent checklist. For model evaluation discipline, reuse the fixed-fixture approach from the Grok 4.5 coding-agent evaluation.
What changed, and where it is available
xAI's July 31 release separates consumer and API availability:
| Surface | Confirmed boundary |
|---|---|
| grok.com and iOS | Image and voice references started in the US for SuperGrok Heavy and Plus, with other tiers rolling out over the following days |
| Android | Text-to-video and native 1080p are generally available; the release does not promise the same immediate reference rollout |
| xAI API | Image references, text-to-video, and native 1080p are described as live on grok-imagine-video-1.5
|
| Voice reference API | Available on request, not an automatic entitlement |
The current model page lists text and image inputs, video output, the aliases grok-imagine-video-1.5-preview and grok-imagine-video-1.5-2026-05-30, and availability in us-east-1 and us-west-2. Check the model list for your team rather than assuming every region or key has identical access.
Audit the API contract before coding around it
The release is newer than several linked documentation pages. Treat the differences as a test requirement, not as permission to choose whichever statement is convenient.
| Capability | New release says | Older/general docs say | Rollout decision |
|---|---|---|---|
| Text-to-video | Live for 1.5 | The current model page now lists text input | Run one prompt-only job and record the returned model |
| Image-to-video | Supported | Supported | Use as the baseline canary |
| Reference-to-video | Up to seven references; API example uses reference_image_urls
|
July 27 guide says 1.5 is unsupported and shows reference_images for the base model |
Start with the newer SDK example; fail closed on invalid_argument
|
| Native 1080p | Release says text-to-video and image-to-video | General guide still says 1080p is image-to-video only | Test prompt-only and image-input jobs separately |
| Voice references | Consumer rollout; API access on request | No general self-serve contract | Keep the feature disabled until the account is approved |
| Mixed modes | Not promised |
image plus reference_images returns 400
|
Keep one mode per request |
This mismatch is likely documentation propagation, but that is an inference. Only a successful request from your target account proves current availability for that account.
Calculate the full clip cost
xAI's pricing page lists a per-image input charge plus per-second video output:
| Resolution | Output price | 6-second clip + one image | 6-second clip + seven images |
|---|---|---|---|
| 480p | $0.08/sec | $0.49 | $0.55 |
| 720p | $0.14/sec | $0.85 | $0.91 |
| 1080p | $0.25/sec | $1.51 | $1.57 |
The formula is:
estimated_cost = duration_seconds * resolution_rate + reference_image_count * $0.01
Do not multiply expensive retries blindly. A three-variant, six-second 1080p test with seven references costs about $4.71 before any failed or repeated generations. Put a job-level limit and daily budget around the worker; the API hard-spend-limit runbook provides a reusable control pattern even though the provider differs.
Run a bounded reference-image canary
1. Freeze one six-second fixture
Use a non-sensitive product object, one cleared background, and a short motion prompt. Keep duration, aspect ratio, seed-like inputs, and evaluation rubric fixed. Do not begin with a real person's face or voice.
2. Send one release-shaped SDK request
The July 31 release uses this Python shape. It is a contract probe, not proof that every account already has the feature:
import os
import xai_sdk
client = xai_sdk.Client(api_key=os.environ["XAI_API_KEY"])
result = client.video.generate(
model="grok-imagine-video-1.5",
prompt="Slow push-in; preserve the product shape and label.",
reference_image_urls=["https://example.com/cleared-product.jpg"],
duration=6,
aspect_ratio="16:9",
resolution="720p",
)
print(result.url)
Keep the key in environment configuration. Never log it, embed it in a client app, or place it in a reference URL.
3. Persist the result and evidence
The video API is asynchronous, and returned xAI URLs are temporary. Persist the asset promptly if your data policy permits it. Store the request ID, requested model, mode, reference count, duration, resolution, terminal status, error code, measured latency, estimated cost, and output checksum.
4. Judge identity and scene separately
For every reference, state the single attribute it is meant to hold: product geometry, character appearance, location, clothing, or palette. Mark a generation failed when an unrelated reference leaks into another role, even if the clip looks impressive.
Eight acceptance tests
| Test | Expected result |
|---|---|
| One reference at 720p | Job reaches done; intended object remains recognizable |
| Seven references | Accepted without silent reference loss; each role is evaluated separately |
| Eight references | Rejected before paid fan-out, or blocked by your own validator |
| Prompt-only 1080p | Either succeeds and records 1.5, or disables that route without retry loops |
| Image-to-video 1080p | Succeeds within the documented duration and cost boundary |
Conflicting image and reference mode |
Client validation stops the request before the API returns 400
|
| Temporary URL | Worker downloads once, verifies checksum, and records storage outcome |
| Permission, moderation, or overload error | No secret appears in logs; only a bounded transient retry is allowed |
Promote the feature only when the exact production region, account, SDK version, and request shape pass. Keep the older base model route available until the new route has real success-rate and accepted-output evidence.
Copyable rollout record
grok_video_canary:
checked_at: "2026-08-01T09:00:00+08:00"
model_requested: "grok-imagine-video-1.5"
account_region: ""
mode: "reference-to-video"
reference_count: 1
duration_seconds: 6
resolution: "720p"
expected_cost_usd: 0.85
terminal_status: "done | failed | expired"
returned_model: ""
output_persisted: true
checks:
subject_consistency: "pass"
scene_consistency: "pass"
audio_review: "pass"
rights_confirmed: true
rollout: "hold | canary | enabled"
Common mistakes
Copying the headline into production requirements. A release announcement can precede updates to every model card, SDK, region, and account entitlement.
Combining generation modes. Reference-to-video, image-to-video, edit, and extend have different field contracts. Build a router, not one payload containing every optional field.
Ignoring resolution economics. For a six-second clip, 1080p output costs almost three times 480p before retry multiplication.
Keeping only the temporary URL. A completed job is not a durable asset archive.
Using faces or voices without authority. Obtain consent and usage rights, limit retention, and keep a human publication gate. Technical access does not establish legal permission.
FAQ
Does Grok Imagine Video 1.5 support reference images in the API?
xAI's July 31 release says yes and provides a 1.5 SDK example using reference_image_urls. An older reference-to-video guide still says otherwise, so verify the capability with your target account before making it a required production path.
How many reference images can I use?
The release and reference guide describe a maximum of seven. Add a client-side maximum so an eighth image never creates a paid or confusing request.
How much does a 10-second 1080p clip cost?
The listed output price is $0.25 per second, so output is $2.50. Add $0.01 for each input image. Availability for the exact mode still needs an account-level canary.
Is voice reference access automatic in the API?
No. xAI says API voice references are available on request. Keep that route disabled until access is explicitly confirmed.
Top comments (0)