Grok Imagine Image 2.0 API

Grok Imagine Image 2.0 is the second generation of SpaceXAI's Imagine image model, shipped as the Quality Mode inside Grok's image generator on August 7, 2026 and exposed over API on August 12. It handles two jobs in one endpoint: generating images from a text prompt, and editing existing images from up to 5 references.

The training split is unusual and worth knowing before you pick it. Rather than one general aesthetic pass, SpaceXAI trained the model separately on photography, graphic design, and illustration. That shows up in three measurable places: lighting and material response on product shots, layout planning on multi-element compositions, and small-point text that survives rendering. It is the reason the model indexes higher on the edit board (1,439 Elo) than on pure text-to-image (1,316 Elo) — it is a stronger editor than it is a from-scratch generator.

Note the provider name. xAI was absorbed into SpaceX, and this model is listed on public leaderboards under the SpaceXAI brand. If you are searching for a "SpaceXAI Grok Imagine Image 2.0 API" and an "xAI" one, they are the same model.

All four models are live on GPTProto under the same key, so this comparison is about fit, not access. Scores below are LMArena Elo from the August 25, 2026 snapshot.

Read the table by job, not by total score.

If the work is editing — retouching a product shot, swapping a background, holding one character across a series — Grok Imagine Image 2.0 is the strongest pick on this list after GPT-Image-2, and the 24-point Elo gap between them (1,439 vs 1,463) is narrower than the gap in operating cost predictability. At a flat $0.04 per image, a 1,000-image edit batch costs $40 regardless of resolution or prompt length. GPT Image 2 bills per token, so the same batch lands somewhere in a range you only know after the fact — fine for exploratory work, awkward for a quoted client job.

If the work is text-to-image from scratch, GPT Image 2 wins on raw preference (1,382 Elo) and there is no honest way to argue otherwise. Grok Imagine Image 2.0 sits at 1,316, ahead of Seedream 5.0 Pro and Nano Banana 2, but behind OpenAI.

If the deliverable carries dense in-image copy in a non-Latin script, Seedream 5.0 Pro's 14-language text rendering is the more direct tool, and its native 2K comes in at $0.081 — double Grok's 2K cost.

If throughput matters more than the last few Elo points, Nano Banana 2 at $0.0402 is effectively the same price as Grok Imagine Image 2.0 and generates faster, but caps at 1K and offers no region-mask equivalent to Magic Wand.

One caveat on the numbers: Grok Imagine Image 2.0's Arena scores carry wider confidence intervals (±12 on text-to-image) than the established models, because it has 2,675 votes against GPT-Image-2's 75,014. Treat it as a strong first-tier signal rather than a settled ranking.

Switching from the SpaceXAI API

If you already call SpaceXAI directly, the migration is a base_url swap and a model-string swap — the request body stays OpenAI-compatible. Three differences to plan for:

Model naming. SpaceXAI's own API exposes this generation through its quality-tier model names. On GPTProto the string is grok-imagine-image-2.0, one string covering both the low and medium quality tiers via the quality parameter rather than two separate model IDs.

Account surface. Direct access requires a SpaceXAI account and its own prepaid balance. On GPTProto the same key and the same balance also reach Seedream 5.0 Pro, Nano Banana 2, GPT Image 2, and 200+ other models, so an A/B test across four image models is a one-line model-string change instead of four sign-ups, four billing relationships, and four sets of rate limits to reconcile.

Fallback path. Because the models share one balance, a failed or filtered generation can retry against a different image model in the same code path. Keep a secondary model string in config — gemini-3.1-flash-image is the closest match on price and modality at $0.0402 per image.

Content Policy and Use

Grok Imagine Image 2.0 is positioned by SpaceXAI for commercial and professional work, a deliberate repositioning after the January 2026 deepfake incident that led xAI to restrict image generation to paid tiers. Two practical consequences for API users.

First, the model ships with pre-configured templates aimed at commercial categories — product stills, professional headshots, e-commerce listings, game concept art, marketing posters. These are product-surface presets rather than API parameters, but they signal what the training was optimized toward, and prompts written in those registers land more reliably.

Second, requests are subject to SpaceXAI's upstream content filtering, which GPTProto passes through unchanged. GPTProto does not add a second moderation layer and does not remove the provider's. A filtered request is a provider decision, not a platform one. If your pipeline needs a fallback for filtered generations, route the retry to another image model on the same key rather than reworking the prompt in place.