Grok Imagine Image 2.0 Jumps to No. 4 in Artificial Analysis

xAI’s Grok Imagine Image 2.0 has reached No. 4 on the Artificial Analysis text-to-image leaderboard, making it the highest-ranked non-OpenAI model in that evaluation. The model entry reports an Elo score of 1154 and places it on the quality-versus-price Pareto frontier, where no evaluated model is both higher-scoring and cheaper.

Artificial Analysis calculates Elo from pairwise image preferences in its Image Arena. The score measures relative preference within the tested field rather than absolute image quality, and rankings can change as new votes and models arrive.

Fourteen places in one generation

The previous grok-imagine-image-quality model ranked No. 18 for text-to-image generation and No. 16 for image editing. Its successor rose 14 places in generation and six places in editing.

At publication, OpenAI’s GPT Image 2.5 Flare (max), GPT Image 2.5 Sunburst (max), and GPT Image 2 (high) occupied the top three text-to-image positions. Grok Imagine Image 2.0 therefore offers the nearest ranked alternative from another provider.

Fourth place, $60 per 1,000

Output pricing starts at $0.04 per image for low quality at 1K resolution. Medium-quality output costs $0.06 at 1K and $0.08 at 2K. Editing also adds $0.01 for each input image.

A 1,000-image batch at medium-quality 1K resolution costs about $60 in output fees. Comparable OpenAI models immediately above it cost roughly $210 per 1,000 images, putting xAI’s price near 29% of that amount. This combination of score and cost places the model between Microsoft’s MAI-Image-2.6 and OpenAI’s GPT Image 2.5 Flare on the leaderboard’s quality-versus-price curve.

Strong typography, slow generation

Artificial Analysis separates individual capabilities from broader use cases. Grok Imagine Image 2.0 scores particularly well in these capability groups:

• Knowledge: real landmarks, species, and domain-specific facts
• Text rendering: long passages, small type, symbols, and stylized lettering
• Reasoning: spatial relationships, logical constraints, concept mixing, and idiom interpretation
• Materials: convincing surfaces, textures, and physical properties
• Complex compositions: scenes containing multiple subjects and layout constraints

Its closest use-case results are in animation and gaming, followed by productivity and knowledge work such as diagrams, infographics, and slides. Consumer illustration and UI/UX design also rank among its stronger workloads.

Image-editing results favor scene and style changes, including relighting, restyling, and background replacement. Identity-preserving edits and enhancement or restoration also perform well relative to its other editing categories.

Median generation time remains the main operational constraint. Grok Imagine Image 2.0 takes 84 seconds to produce a 1024-by-1024 image at medium quality, according to the leaderboard measurements.

Compared with its five-second predecessor, the new model exchanges substantial speed for higher preference scores. Its 84-second median still beats the two listed OpenAI variants, while MAI-Image-2.6 and Nano Banana 2 suit workflows that require faster responses.

Editing gains finer controls

xAI expanded the editing system beyond general prompt-based changes. The developer documentation lists several production-oriented additions:

• Region-level editing: changes can target a selected part of an image.
• Background removal: the model can isolate a subject without a separate segmentation tool.
• Multi-reference composition: one request can include as many as five reference images.
• Typography and layout: improved planning helps preserve small text and dense visual structures.
• Iterative preservation: supplied subjects and other requested elements remain more consistent across successive edits.

Multi-reference composition can combine inputs such as a product photograph, style reference, and layout template in one request. At the listed editing rate, five source images add $0.05 before output charges.

API defaults can change the result

The model is available from the xAI API under grok-imagine-image-2.0, through Grok’s apps and image-generation site, and from third-party gateways including OpenRouter, fal, and Replicate. Third-party pricing, parameters, and defaults may differ from xAI’s direct API.

The xAI Python SDK accepts the model identifier through client.image.sample. This example fixes the quality level so deployment behavior does not depend on the current default:

import xai_sdk

client = xai_sdk.Client()

response = client.image.sample(
    prompt=(
        "A concert poster for a synthwave band, "
        "bold retro typography, sharp small print"
    ),
    model="grok-imagine-image-2.0",
    quality="medium",
)

print(response.url)

xAI changed the default quality from medium to auto shortly after launch. When callers omit the parameter, auto currently resolves to low for text-to-image generation and medium for image editing. Production code should set quality explicitly when cost, latency, and output consistency need predictable bounds.

Where the model fits

• Strong candidates: character design, concept art, game assets, comics, posters, diagrams, presentations, and UI mockups.
• Best operating pattern: asynchronous jobs, batch generation, and workflows that can tolerate an 84-second median response.
• Faster alternatives: Nano Banana 2 and MAI-Image-2.6 for interactive generation, or the older xAI model when its lower quality is acceptable.
• Cost-sensitive editing: multi-image composition and targeted edits, provided the additional per-input fee is included in estimates.

Elo aggregates preferences across many prompts, so production tests should use representative content, aspect ratios, typography, reference-image counts, and editing sequences. For workloads that reproduce its benchmark strengths, Grok Imagine Image 2.0 offers the highest-ranked non-OpenAI result at a substantially lower unit price than the three models above it, with latency as the primary trade-off.