Grok Imagine 2.0 vs GPT Image 2: Scores, Editing, and API

xAI shipped Imagine Image 2.0 on 7 August 2026, and the Arena boards it cited put the model second in the world in both text-to-image and image editing. First place in both went to OpenAI's GPT Image 2^[1]. That is the headline, and it is more interesting than it sounds, because in a Grok Imagine 2.0 vs GPT Image 2 comparison the gap is not where most people assume it is.

If you are weighing Grok Imagine 2.0 vs GPT Image 2 to decide what to build on, the ranking is the least useful number on the page. What matters is that one of them is callable from an API today and the other is not, and that the two models are optimised for different halves of the same job.

What the Grok Imagine 2.0 Arena scores actually say

On the leaderboards xAI referenced on announcement day, Grok Imagine 2.0 posted 1320 (±12) in Text-to-Image and 1439 (±8) in Image Editing^[1]. The previous generation, Grok Imagine Quality, sits at 1228 and 1390 on those same current boards.

That works out to roughly +92 Elo on generation and +49 on editing. It is not a rounding error, and it is not "the skin looks a bit better" either. A jump that size usually means the model changed how it handles instructions, not just how it renders pixels.

Here is the asterisk. Grok Imagine 2.0's scores still carry a Preliminary label because the vote count is low. "Strong early placement" is the honest reading. "Second best image model in the world, settled" is not, at least not yet. Any Grok Imagine 2.0 vs GPT Image 2 verdict written this month is provisional by definition.

The part that trips people up

Grok Imagine 2.0 has no public API today. xAI's announcement ends with the words "API access is coming soon" and stops there^[1]. No endpoint, no parameter reference, no date.

Meanwhile there is a Grok image model on the API today, and it is not this one. grok-imagine-image-quality has been callable since May 2026^[2]. It is the previous generation, the 1228/1390 row in the table above. Wire up an integration this week expecting Grok Imagine 2.0 behaviour and you will get Quality behaviour, then spend an afternoon wondering why the editing feels weaker than the demos.

GPT Image 2, by contrast, is callable now. So the practical question for the next few weeks is not which model is better. It is which model you can actually call.

Where Grok Imagine 2.0 is genuinely stronger

The two models are not competing for the same slot in a pipeline.

GPT Image 2 wins on one-shot correctness. Complex instruction following, text rendering, subject preservation, and fewer wasted retries. Through the Responses API you can keep a conversation going and make high-fidelity edits against it^[3]. If your workflow is "get one image exactly right, then revise it precisely", that is the stronger tool, and its Arena placement backs that up.

Grok Imagine 2.0 wins on the loop. The Grok Imagine 2.0 editing model is the interesting half: a magic wand that changes only the region you point at, segmentation for precise areas, background removal that exports subjects on transparency, and multi-reference editing that accepts up to five inputs in a single generation.

That last capability is the underrated one. Instead of loading a single reference image with every job at once, character, wardrobe, location, lighting and overall look, you split them across five inputs and control exactly what gets inherited from where.

The older failure mode is what this fixes. You generate a good frame, one hand comes out wrong, you regenerate to fix it, and now the face, the body, the lighting and the composition have all drifted too. Region-scoped editing means the parts you already approved survive the fix. In practice that turns image generation into something closer to a photo shoot with a retouching pass, rather than pulling a slot machine lever until a whole frame lands.

Smart Resize belongs in the same category. Nine published aspect ratios run from 1:2 through 2:1, and rather than cropping an existing composition the model fills the frame for the ratio you pick. A 2:3 portrait becomes a 9:16 vertical without losing what sat outside the new crop.

The video adjacency nobody counts

GPT Image 2 does not output video. Grok Imagine does, and the two halves are built to hand off to each other. That adjacency rarely shows up in a Grok Imagine 2.0 vs GPT Image 2 feature table, and for some teams it decides the whole question.

Video 1.5 covers text-to-video, image-to-video, reference-to-video with up to seven images, native 1080p, clips up to 15 seconds, and audio generated in the same pass. A 6-second 720p Fast generation dropped from over 40 seconds to roughly 25^[4]. On the Image-to-Video Arena board it sat third as of 2 August 2026, inside a top group whose error bars overlap.

For a still image, a few seconds of motion is often enough. A gaze that shifts. Shoulders that rise with a breath. Hair moving in wind. A slow push-in. If your output lands on social, that hand-off from image to short clip is worth more than 90 Elo on a leaderboard.

How this fits next to Gemini and Krea

Four models, four different jobs:

Model	What it is best at
GPT Image 2	Landing one image exactly as instructed, then editing it precisely
Gemini / Nano Banana 2	World knowledge, long context, conversational multi-modal editing
Krea 2	Exploring an aesthetic direction rather than nailing one correct answer
Grok Imagine 2.0	Generating, region-fixing and animating fast in one loop

Gemini 3.1 Flash Image accepts audio and video as input, can fact-check through search, and generates up to 4K. If you need a model to reason about a complex scene or hold continuity across a long conversation, that is where it pulls ahead.

Krea 2 is built around exploration rather than execution. Style references, moodboards, intensity and complexity sliders, LoRAs. It suits finding a look across a series more than executing a fixed brief.

None of that is a ranking. It is four different shapes of work, and picking by leaderboard position alone is how teams end up with the wrong tool.

What we do not know yet

Worth stating plainly, because a lot of coverage is guessing.

xAI described the older Aurora model as an autoregressive image generation model^[5]. For Grok Imagine 2.0, the architecture, parameter count, training method and any MoE structure are simply not published. Claims that it is "the same autoregressive MoE as Aurora", or that it uses a particular DiT or VAE, are inference dressed as fact.

Pricing is the same story. Nothing published, for either the model or the forthcoming API. Anyone quoting a per-image rate for Grok Imagine 2.0 today invented it. Output resolution ceilings and rate limits are undisclosed as well.

Picking a model for this month

If you need an image model in production right now, GPT Image 2 is the one with an endpoint, and it is first on both boards it was measured against. That is an easy call, and no amount of preliminary Elo changes it.

If your work looks like a loop, generate, fix one region, keep the subject consistent across a set, then push the best frame into a short clip, the Grok Imagine 2.0 vs GPT Image 2 decision stops being a quality question and becomes a workflow one. The practical move is to build against a shared image endpoint today with a model that ships, then swap the model id when Grok Imagine 2.0 opens up. On reAPI that is a one-line change rather than a migration, and the same API key already works across every model on the platform.

What the editing model changes in practice

Most image APIs treat editing as a second endpoint that happens to accept an image. Grok Imagine 2.0 treats it as the same system that generated the frame, and that distinction shows up in the failure modes rather than the demos.

Take a product shot where the label reads wrong. On a generate-only model you rewrite the prompt and roll again, and the lighting shifts, the shadow moves, the reflection changes, and now you are comparing two images that differ in five ways instead of one. With region-scoped editing you point at the label. Everything else is byte-for-byte what you already signed off on.

The same logic covers a series. Give Grok Imagine 2.0 a character reference, a location reference and a wardrobe reference as separate inputs, and you can run the same subject through morning, afternoon and interior setups without the face drifting between shots. One reference image carrying all of that at once is where consistency usually breaks.

Background removal fits the same workflow. Exporting a subject on transparency is not a headline capability, but it is the step that decides whether an output goes straight into a layout or needs a round trip through a mask tool first.

None of this makes Grok Imagine 2.0 a better model than GPT Image 2 on raw instruction following. The Arena numbers say otherwise, and preliminary or not, they point the same direction. What it changes is how many generations you burn to get from a decent frame to a finished asset, and that cost never appears on a leaderboard.

References

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
OpenAI. Image generation guide. Retrieved August 2026 from platform.openai.com/docs/guides/image-generation
xAI. Grok Imagine Video 1.5. Retrieved August 2026 from x.ai/news/grok-imagine-video-1-5
xAI. Grok Imagine API. Retrieved August 2026 from x.ai/news/grok-imagine-api