AI&Design
All change articles

Grok · Announcement · Official source

Grok Imagine’s legacy Quality Mode API

On May 6, 2026, xAI announced Quality Mode for enterprise image generation and editing through the Grok Imagine API. The launch article’s Python sample calls the model as grok-imagine-image-quality and returns the image URL. xAI showed examples of photographic scenes and text in images, and promoted the model for tasks where realism and text rendering matter. This was a named image model that teams could call from their own code. The announcement also shows stills as inputs for separate video work. xAI’s May 6 announcement

The names overlapped across three different changes. On May 15, xAI retired the earlier grok-imagine-image-pro model and redirected that slug to the May Quality model. On August 7, xAI announced Imagine Image 2.0 as the new consumer Quality Mode on the web, iOS, and Android. Its developer model ID is grok-imagine-image-2.0. The legacy enterprise model continued to be selectable in OmniaKey’s October test, alongside the base model and 2.0. May 15 migration guide · August 7 Image 2.0 announcement · OmniaKey comparison

On July 3, fal published John Ozuysal’s edit comparison. He used the same skincare product image and edit brief across models. The brief asked for a sage background, a brushed-gold dropper cap, and a serif tagline while retaining the bottle, linen, and eucalyptus. On fal’s legacy Quality edit endpoint, Ozuysal reported sharper detail and cleaner tagline lettering than in the standard Grok edit, with incorrect wording still visible on the bottle. fal shows the resulting image and sells access to the endpoint. The published example records one operator’s assessment of a particular brief. fal’s test and output

OmniaKey COO Kiki published an October 18 comparison the next day. The team ran four prompts once per model through OmniaKey with no specified size or aspect ratio. The legacy Quality poster correctly spelled the two requested lines and added an invented shop name and street address. The same model followed the earbud product brief and, in Kiki’s judgment, produced the cleanest overhead composition in the fruit test. All three models got the fruit counts right. OmniaKey sells model access, and its Grok workflow in the article covers generation. The article’s separate timed Python exercise calls Image 2.0 to render a bicycle. These observations give examples of typography and composition from individual generation jobs. OmniaKey method and outputs

On November 2, xAI retired the legacy model. Requests to its Quality and Pro slugs began serving Image 2.0 at Low quality on xAI’s generation and edit endpoints. The migration guide says both URLs and request shapes continue to work, and gRPC responses identify the replacement as grok-imagine-image-2.0-low. Output pricing fell one cent per image at both supported resolutions. On November 5, a Reddit poster said the old model gave better results for their images and asked xAI to release its weights. The submission contains a preference and no comparative images or prompts. November 2 migration guide · November 5 discussion

The April 29–May 20 public Reddit search yielded zero verified submissions tying a workflow to the legacy enterprise API. One candidate concerns consumer quota counters, and another describes repetitive images after returning to a subscription, with the Quality label appearing in the replies. Genflick’s nearby May 2 operator benchmark names the base API model in its method. The earliest retained exact workflow is fal’s July edit example. The October comparison adds a generation example with the old model explicitly identified. Both are vendor-published tests with limited samples. Consumer quota thread · Image repetition thread · Genflick benchmark method