xAI Imagine API Update: Auto Quality & Pricing Impact

xAI has updated the Grok Imagine image API in three places: automatic quality selection, multi-image editing, and wide-format output. The change applies to grok-imagine-image-2.0 and is live in xAI’s August release notes.

The most important pricing change is indirect. xAI did not cut every Imagine rate; it changed what happens when developers omit quality. A generation request that previously defaulted to medium now defaults to auto, which currently serves low. Editing through auto currently serves medium.

What changed

API behavior	Before	Now	Practical effect
Omitted quality	medium	auto	Generation currently routes to low; editing routes to medium
Image-edit references	Up to 3	Up to 5	More products, characters, or style references in one request
Wide aspect ratios	Existing ratio set	Adds 21:9 and 5:2	Native cinematic frames and banner assets

auto is not a permanent synonym for one fixed tier. xAI describes it as service-selected quality and says images are billed at the quality actually served. Developers can still pass low or medium to pin the behavior.

Pricing impact: the generation default gets cheaper

For the default 1k resolution, moving an omitted-quality generation from medium to auto’s current low tier reduces the served-image charge by one-third. At 2k, the reduction is one-quarter. Explicit medium requests keep the medium rate.

Imagine 2.0 generation	Previous omitted-quality cost	Current omitted-quality cost	Change
1k image	$0.06	$0.04	33% lower
2k image	$0.08	$0.06	25% lower

These are xAI’s public model-configuration rates checked August 31, 2026. The simplified xAI pricing page lists Imagine 2.0 from $0.04 per image, while the model configuration distinguishes resolution and quality. Editing stays on medium under today’s auto behavior, so teams should not apply the generation savings to edits.

At 10,000 successful 1k generations, today’s auto routing would reduce the image line item from $600 to $400 compared with the previous medium default. Real campaign cost also includes rejected outputs, retries, storage, moderation failures, and any surrounding text-model or tool usage.

Who benefits—and who should be careful

High-volume thumbnail, concept, advertising, and catalog pipelines benefit most from the lower generation default. The new 5:2 ratio can produce wide display banners without cropping, while 21:9 targets cinematic storyboards and headers. Five-reference editing is useful when an output must combine a product, person, background, brand style, and layout cue.

The trade-off is silent behavior change. A deployed client that omitted quality may now receive low instead of medium generation quality. That can save money, but it can also change fine detail, text rendering, or approval rates. If more outputs are rejected, the lower per-image charge may not lower cost per approved asset.

The five-reference limit also does not guarantee that every reference receives equal weight. Test identity, product accuracy, composition, and instruction following with production-shaped inputs before expanding a workflow.

Labs status: image benchmark blocked by the current harness

Grok Imagine Image 2.0 is not inserted into the current Cost-per-Task leaderboard. That suite grades deterministic text answers and token cost; it cannot honestly score visual quality, edit fidelity, or cost per accepted asset. A valid image study needs pinned model and quality settings, fixed generation and editing prompts, licensed reference images, repeated samples, blind human acceptance labels, latency, retries, safety failures, the returned billed cost, and a disclosed cost-per-accepted-image rule.

What API teams should do now

1. Search generation and editing calls for requests that omit quality.
2. Pass low for cost-controlled drafts or medium for a stable higher-quality path; use auto only when service-selected routing is acceptable.
3. Log the returned model, resolution, quality where available, and cost_in_usd_ticks for every request.
4. Re-run the same acceptance set at 1k and 2k; compare cost per approved image, not only cost per call.
5. Add 21:9 and 5:2 to validation schemas before exposing them in a UI.
6. Test four- and five-reference edits separately from the old three-image path and set an explicit output aspect ratio when the first input should not control it.

For current xAI text, voice, and video rates, use our xAI API pricing page. Compare the surrounding reasoning layer with OpenAI pricing and use the AI token calculator for token-metered steps; image jobs must be budgeted per served image. Our AI image generation pricing guide covers the wider market.

Teams that want a managed image-model control group can check Novita’s current image catalog, then compare the same prompts and approval criteria. Confirm the exact model and live rate before treating it as equivalent.

Affiliate disclosure: we may earn a commission from the sponsored link above. It does not affect this analysis.

Bottom line

xAI’s Imagine update lowers the current default cost of image generation when quality is omitted, but it also changes output behavior. Pin the quality for predictable production, keep auto for flexible routing, and measure approval rate before counting the headline saving. Five reference images and two new wide ratios broaden the API without requiring a new model ID.

Sources: xAI’s official release notes, image generation guide, multi-image editing guide, and Imagine pricing. Pricing and API behavior verified August 31, 2026.