SpaceXAI is retiring a model slug on 2 November, and the replacement will not draw the same picture.
From that date, requests to grok-imagine-image-quality will be served by grok-imagine-image-2.0 at quality: low. The request and response shapes do not change, so nothing in a caller’s code breaks. The images will simply come back different.
What actually changes
The company is explicit that no code change is needed and equally explicit that output will differ. Anyone with a pipeline tuned around the old model’s characteristics — a house style, a prompt library, a set of accepted results — has until 2 November to test 2.0 at low and find out what shifted.
Two months of notice for a silent output change is generous by the standards of this industry. It is still a change that will surprise anyone who reads the migration note as a pricing update and stops there.
The price moves in the customer’s favour: 2.0 at quality: low costs one cent less per image than the slug it replaces. grok-imagine-image, the 1.0 model, is unaffected.
Auto is now the default
Separately, the quality parameter has gained an auto setting, and auto is now the default rather than medium. It picks a tier per request to cut latency: today that means low for generation and medium for editing.
The billing line is the part worth reading twice. You are billed at the tier that was served, not the tier you asked for. A default that chooses its own price point is convenient until a workload shifts toward editing and the invoice moves with it.
More references, wider frames
Editing now accepts up to five reference images per request, up from three. SpaceXAI’s own example is one character, three props and one location, composed into a single prompted scene — which is less an incremental limit change than a different way of working.
Generation and editing also gained two aspect ratios, 21:9 and 5:2, aimed at widescreen and banner formats.
The benchmark caveat
SpaceXAI publishes an Arena Single Image Edit radar chart putting 2.0 at quality: low ahead of grok-imagine-image-quality in every category, with an overall score of 1,439 and category scores from 1,432 for photorealistic to 1,474 for text rendering.
That comparison is the company’s own presentation of a third-party leaderboard, selected to show the replacement beating the thing being replaced. The underlying Arena rankings are independent; the framing is not. When SpaceXAI announced 2.0 on 7 August it placed second globally on both the Arena text-to-image and image-edit boards, behind OpenAI’s GPT-Image-2.
What to watch
Whether “output will differ” turns out to mean a refinement or a re-tune. The company has not published a side-by-side of the old slug against its replacement on the same prompts, which is the comparison a migrating developer actually needs.
