GPT Image 2 leads in text, realism; Nano Banana 2 wins anime

OpenAI released GPT Image 2 in late April. The model runs on a GPT-5.4 backbone with native reasoning built into its architecture. OpenAI retired DALL·E 3 and GPT Image 1.5 and scheduled their shutdown for May 12. GPT Image 2 posted a 242-point lead on the Image Arena leaderboard over the next model on that ranking.

Independent tests compared GPT Image 2 and Google’s Nano Banana 2 across seven categories: photorealism, classical painting, anime-style illustration, lettering and signature design, aerial spatial composition, high-density text recall, and agent-enabled information design. The same seven-category framework used in prior comparisons guided the evaluation.

OpenAI reported that GPT Image 2 achieves roughly 99% character-level accuracy across Latin, CJK, Hindi and Bengali scripts. The model supports up to 4K resolution and can generate up to eight coherent images from a single prompt while keeping characters and objects consistent across the batch. Access is tiered: Instant Mode is available to all ChatGPT users, including free accounts, while Thinking Mode, which lets the model research, plan and self-check before generating, requires Plus, Pro or Business subscriptions. The official API was slated to open in early May. Direct generation through ChatGPT or third-party proxies ran at about $0.01–$0.03 per image. OpenAI’s token-based pricing was set at $8 per million input tokens and $30 per million output image tokens; Nano Banana 2’s output token price was noted at $60 per million at similar resolution tiers.

Test results show GPT Image 2 leading on photoreal texture, precise text rendering and faithful edits. In a cinematic portrait test with constraints on clothing, props and lighting, GPT Image 2 produced skin with natural subsurface scattering, accurate garment texture and the specified blueprint held in the right hand. Nano Banana 2 produced a portrait with a more natural gaze and different color grading and held different blueprints than requested.

In a Rembrandt-style oil painting prompt with multiple light sources and detailed props, GPT Image 2 reproduced color temperatures and material interactions consistent with oil painting, but reviewers reported oversharpening and artifacts when many constraints were applied. Nano Banana 2 returned an image closer to high-fantasy illustration rather than an oil painting, with some genre cues missing.

Nano Banana 2 outperformed GPT Image 2 in anime-style illustration. A key-visual prompt that called for cel shading, varied ink outlines, legible kanji on talismans and a specific twilight gradient produced a result on Nano Banana 2 that matched the technical requirements, including readable ofuda characters and clear tail rendering on a multi-tailed creature. GPT Image 2 returned an anime-style image that missed some subsurface glow and produced inconsistent tail representations.

On a demanding aerial steampunk composition that required distinct depth planes and multiple readable text elements, Nano Banana 2 yielded clearer geometry and atmospheric layering. GPT Image 2 reproduced specified text elements more accurately in that test but showed partial mid-ground collapse in depth.

In a dense-lettering urban night scene that required many readable text elements across varied surfaces, GPT Image 2 reproduced near-complete signage, posters and small copy with accurate lighting and textures. Nano Banana 2 produced images that reviewers found more visually pleasing in some cases but that missed specific text treatments and poster details.

The models showed different workflow strengths. GPT Image 2’s batch consistency and high text fidelity align with production tasks such as multi-format children’s book art or coordinated campaign assets. Nano Banana 2 performed well for stylized illustration, rapid iterations and complex spatial scenes when guided by targeted prompting techniques. Reported failure modes include oversharpening and artifacts for GPT Image 2 on long, highly specific prompts, and repeated prior outputs or loss of fine text detail for Nano Banana 2.

OpenAI released GPT-5.5 for tasks that require the model to research, plan and act with reduced human supervision and rolled that update out to paid ChatGPT tiers. The GPT Image 2 API opening was scheduled for early May and DALL·E 3 and GPT Image 1.5 were retired on May 12. Buyers and developers will evaluate text fidelity, photorealism, stylized art quality and pricing when choosing between the two models.