Which Grok Imagine Model to Use: 1.0 vs 1.5 vs Image 2.0

Three Grok Imagine models have shipped in four months, and picking between them is not a question of which is newest. Each Grok Imagine model does a different job, one of them cannot be called from an API at all, and the naming makes the whole thing harder than it needs to be.

Here is the short version. Grok Imagine 1.0 Video and Video 1.5 generate video. Imagine Image 2.0 generates images. The version numbers run across two different modalities, so "2.0 is the upgrade to 1.5" is wrong in the way that matters.

What each Grok Imagine model actually is

Release	Date	Modality	On reAPI
Grok Imagine 1.0 Video	7 May 2026	Video	Live
Grok Imagine Video 1.5	4 June 2026	Video	Live
Imagine Image 2.0	7 Aug 2026	Image	Coming soon

The trap is in the middle column. Someone shopping for a Grok Imagine model sees 2.0 as the flagship and assumes it supersedes 1.5. It does not. If you need video, 1.5 is the current generation and 2.0 has nothing to offer you. If you need images, 2.0 is the one you want and it is not callable yet.

Video 1.5 is the working video model

Video 1.5 covers text-to-video, image-to-video, and reference-to-video with up to seven reference images. It generates native 1080p, clips up to 15 seconds, and audio in the same pass as the picture, which means ambience, effects and dialogue arrive already synced instead of being layered in afterwards.

Speed moved too. A 6-second 720p Fast generation went from over 40 seconds to roughly 25^[1]. On the Image-to-Video Arena board it placed third as of 2 August 2026, inside a leading group whose error bars overlap, so treat the ordering inside that group as noise.

Practically, this is the Grok Imagine model you reach for when a still image needs to become a few seconds of motion. Not a performance, just a gaze that shifts, hair that moves, a slow push-in.

What 1.0 Video is still for

Grok Imagine 1.0 Video predates 1.5 by a month and remains live on reAPI. It is the cheaper tier rather than a deprecated one. If your workload is high volume and short, and you do not need 1080p, seven references or the faster turnaround, the older Grok Imagine model still clears the bar.

The honest framing: 1.5 is better at nearly everything, and the reason to run 1.0 is cost per clip, not capability.

That still makes it a real choice in two situations. The first is bulk generation where clips are short, get cut down further, and never leave a vertical feed, so 1080p buys nothing. The second is A/B work where you are testing prompts or concepts rather than producing finals, and burning the cheaper Grok Imagine model through fifty variants beats burning the expensive one through fifteen. Once a concept is locked, re-render the keeper on 1.5.

Image 2.0 is a different product with a version number

Imagine Image 2.0 shipped on 7 August 2026 as the new Quality Mode inside Grok^[2]. On the Arena boards xAI cited it placed second in the world in both text-to-image and image editing, behind OpenAI's GPT Image 2 in both. Those scores still carry a Preliminary label because the vote count is low.

What separates it from a plain text-to-image model is that editing is built in rather than bolted on. A magic wand changes only the region you point at. Segmentation selects precise areas. Background removal exports subjects on transparency. Multi-reference editing accepts up to five inputs in a single generation, so a character, a location and a prop can be supplied separately instead of crammed into one reference image.

There is one catch, and it is the thing most people get wrong. Image 2.0 has no public API. xAI's announcement ends with "API access is coming soon" and gives no endpoint and no date.

The naming collision that costs people an afternoon

There is already a Grok image model on the API, and it is not 2.0. grok-imagine-image-quality has been callable since May 2026^[3]. It is the previous generation, scoring 1228 and 1390 on the same boards where Image 2.0 posted 1320 and 1439.

So if you integrate a Grok Imagine model for images this week, you get Quality, not 2.0. The region-scoped editing and five-reference compositing that make 2.0 interesting are not in that endpoint. Budget for the older behaviour or wait.

Picking a Grok Imagine model by job

Work backwards from the output, not the version number.

Video, shipping today. Video 1.5. Native 1080p, up to 15 seconds, audio in the same pass, seven reference images.

Video, cost-sensitive at volume. 1.0 Video. Older, cheaper, still live.

Images, shipping today. Not a Grok Imagine model. Use something with an endpoint, then swap later.

Images, planning ahead. Image 2.0, when the API opens.

That last row is worth designing for now. Every image model on reAPI speaks the same async request shape, so building against a shipping model today and changing the model id when Image 2.0 lands is a one-line edit rather than a migration. The same API key already works across the platform.

Cost is the part the version numbers hide

Version order implies that newer means pricier, and across this lineup that is not reliable either. On the video side, 1.0 Video and Video 1.5 are billed per second and scale by resolution, so the gap between them widens with clip length rather than staying a flat percentage. A 15-second 1080p render and a 4-second 480p render are different products commercially even though the same Grok Imagine model produced both.

That matters when you are sizing a workload. Teams routinely benchmark on short 480p test clips, pick the newer Grok Imagine model on quality, then discover the real bill lives in the resolution and duration they actually ship at. Run the estimate at production settings before committing.

For Image 2.0 there is no cost conversation to have yet, because no rate has been published. When it lands, the pricing table on its model page fills in automatically, and the credits system does not change: you pay per completed generation, failed generations refund, and credits do not expire.

One more practical note. Because every model on the platform shares one key and one async request shape, comparing two Grok Imagine model options in production is cheap. Point the same integration at each id, run the same prompts, and compare real outputs and real charges rather than leaderboard positions. That is usually a better signal than an Elo delta measured on someone else's prompts.

What is not published

Worth being explicit, because guesses circulate as facts.

Image 2.0's architecture, parameter count and training method are undisclosed. xAI described the older Aurora model as autoregressive, but claims that 2.0 shares that design are inference, not documentation. Pricing for Image 2.0 has not been published either, for the model or the coming API. Any per-image rate quoted for it today is invented.

Where this leaves the lineup

The version numbers suggest a ladder. The reality is two product lines sharing one name: video at 1.0 and 1.5, images at 2.0, plus a fourth model, grok-imagine-image-quality, quietly serving image traffic on the API while 2.0 gets the attention.

So the selection rule is not "take the highest number". Start from the output you need, check whether that Grok Imagine model has an endpoint you can call today, and only then compare quality. Two of the four fail the endpoint test for images right now, and one of them fails it silently by answering with an older generation than you expected.

For video work, the current Grok Imagine model is Video 1.5, and it is live. For image work, the Grok Imagine model worth planning around is Image 2.0, and it is not callable yet. Build on something that ships, keep the model id in config rather than hardcoded, and the day Image 2.0 opens you change one line instead of rewriting an integration.

References

xAI. Grok Imagine Video 1.5. Retrieved August 2026 from x.ai/news/grok-imagine-video-1-5
xAI. Imagine Image 2.0. Retrieved August 2026 from x.ai/news/grok-imagine-image-2
xAI. Grok Imagine Quality Mode API. Retrieved August 2026 from x.ai/news/grok-imagine-quality-mode