Grok Imagine Image 2.0 debuts at No.2 in both arenas, then slips to third

SpaceXAI's new image model entered both Arena leaderboards at No.2 behind GPT-Image-2; by 11 August it had been pushed to third in text-to-image by Microsoft's MAI-Image-2.6, while holding second in image editing.

The result

Arena, the crowdsourced model-comparison platform, published leaderboard placements for the newest image model from SpaceXAI — the company formerly known as xAI, rebranded in July after its merger into SpaceX — on Saturday morning UTC. Grok Imagine Image 2.0, tested in its faster "Low" configuration, entered the Text-to-Image Arena in second place with 1,320 points, behind OpenAI's GPT-Image-2 at 1,380. In the Image Edit Arena it also landed second, scoring 1,439 against GPT-Image-2's 1,463 — a jump from seventh place in that arena for the previous generation, Grok Imagine Image Quality, which Arena ranks 14th in text-to-image.

One of those two placements has already changed. As of 11 August the model sits third in text-to-image with 1,316 points, displaced by the MAI-Image-2.6 preview from Microsoft AI, which entered on Monday at 1,336; GPT-Image-2 leads with 1,381. Ranked below the SpaceXAI entry are Reve AI's Reve 2.1 at 1,302, Meta's Muse-Image at 1,282, Reve AI's older Reve 2.0 at 1,270, Google's Gemini 3.1 Flash Image at 1,264, ByteDance's Seedream 5.0 Pro at 1,258 and Alibaba's Qwen-Image-3.0-Pro at 1,257. In the Image Edit Arena it still holds second, 1,439 against 1,463. Both scores remain provisional: roughly 2,700 text-to-image votes against about 70,000 for the leader, and about 5,900 image-edit votes against 184,000. Positions keep moving as voting accumulates.

What shipped

SpaceXAI released Image 2.0 on 7 August as the "Quality Mode" of Grok Imagine, available on grok.com/imagine and in the company's iOS and Android apps. First-party API access has not opened; the company says it is coming, without giving a date, a model ID or pricing. Developer access so far runs through a third party: a preview of the model has been made available on Vercel's AI Gateway. SpaceXAI describes the model as designed to follow detailed instructions precisely, plan typography and page layout the way a designer would, and hold small text legible inside dense composite visuals.

The release bundles an editing toolkit rather than a bare generator: a magic-wand tool that alters only the region a user points at, segmentation-based selection, background removal with transparent export, multi-reference editing that accepts up to five input images in a single generation, and aspect-ratio conversion via Smart Resize. Templates package workflows for photo retouching, product photography, marketing assets and game art. The company also highlights consistent characters, locations and props across separate generations — a capability it frames as groundwork for later video production.

Why it matters

Human-preference leaderboards have been led by OpenAI's image stack for much of this year, with Reve AI, Microsoft and Chinese entrants such as Qwen-Image-3.0-Pro pressing from below. SpaceXAI took second on both boards at launch using its cheaper inference tier — but within three days the text-to-image order had shifted again, leaving it about 65 Elo points off the top and behind Microsoft's newest preview, while its 24-point editing deficit still stands. The catch for professional users is distribution: with no first-party API and developer access limited to a Vercel gateway preview, the model lives mainly inside Grok's consumer apps, so unlike Qwen's and ByteDance's image and video models — already wired into many third-party platforms — it is only beginning to reach production creative pipelines. Whether the ranking translates into real usage depends on how quickly that API arrives.

← All briefings