GPT-Image-2 First Look: What We Know So Far and How It Compares to Nano Banana Pro

Three anonymous models appeared on LM Arena, stunned testers with near-perfect text rendering, and vanished within hours. The AI image generation landscape is about to shift again.

OpenAI's next-generation image model has been spotted in the wild.

On April 4, 2026, three unidentified models appeared on LM Arena, the popular blind-testing platform for AI models. Within hours, they had left testers stunned with capabilities that clearly surpassed anything currently available from OpenAI, including near-perfect text rendering, eliminated color casts, and dramatically improved world knowledge. Then, just as quickly as they appeared, the models were pulled.

The AI community reached a swift consensus: this was GPT-Image-2, OpenAI's successor to the GPT-Image-1 and 1.5 models that currently power image generation in ChatGPT.

Since then, evidence has continued to mount. As of April 17, the model is being A/B tested within ChatGPT itself. Mobile app strings referencing "GPT-Image-2" have been discovered by developers digging through code updates. And with DALL-E 2 and DALL-E 3 scheduled for retirement on May 12, OpenAI clearly has something ready to fill the gap.

Here's everything we know so far about GPT-Image-2: its capabilities, how it compares to Google's Nano Banana Pro in head-to-head blind tests, where Nano Banana 2 fits into the picture, and what the timeline looks like for a public launch.

How GPT-Image-2 Was Discovered

The story begins with LM Arena, the community-driven platform where AI models compete in blind head-to-head comparisons. Users submit prompts, two anonymous models generate outputs, and users vote on which result they prefer. It's considered one of the most unbiased ways to evaluate AI model quality because testers don't know which model they're judging.

The April 4 Appearance

On the morning of April 4, 2026, three new models appeared on LM Arena under codenames that immediately caught the community's attention:

• maskingtape-alpha
• gaffertape-alpha
• packingtape-alpha

The naming convention alone was a signal. LM Arena codenames are assigned by the platform, not the model providers, but the "tape" theme suggested these were related models, likely variants of the same underlying architecture being tested under different configurations.

What Testers Saw

Within the first few hours of testing, the results were striking. The tape models were generating images with characteristics that no publicly available OpenAI model could match:

• Text rendering that actually worked. UI interfaces with correctly spelled button labels. Watch faces displaying accurate times. Product packaging with readable, properly formatted text. This alone was a massive leap. GPT-Image-1.5, the current production model, manages roughly 90-95% text accuracy. These models appeared to be clearing 99%.
• No yellow color cast. The warm yellow/orange tint that has plagued every version of OpenAI's image generation since DALL-E was simply gone. Colors were neutral, accurate, and true to the prompt descriptions.
• Photorealistic quality at high resolution. The outputs had a level of detail and coherence that suggested a fundamentally different architecture, not just an incremental improvement on the existing model.

The Models Disappeared

Within hours, all three models were removed from LM Arena. This is consistent with how major AI labs typically conduct pre-release testing: deploy briefly to gather real-world performance data, then pull the models before too much information leaks.

It didn't work. Screenshots, comparison images, and detailed analysis had already been shared widely across X (Twitter), Reddit, and AI-focused Discord servers. By the time the models were pulled, hundreds of side-by-side comparisons had been saved, dissected, and debated. The AI community had already reached its verdict: whatever these models were, they represented a generational leap in OpenAI's image generation capabilities.

The codename pattern itself became the subject of speculation. "Maskingtape," "gaffertape," and "packingtape" all reference adhesive tape, a material used to hold things together or seal packages. Some community members interpreted this as a reference to the model "taping together" multiple capabilities (text, image, spatial understanding). Others suggested it was simply OpenAI having fun with codenames. Either way, the tape family had made its mark.

Confirmation Through A/B Testing

As of April 17, 2026, multiple users have reported encountering noticeably different image generation behavior within ChatGPT itself. The symptoms match what was seen on LM Arena: improved text rendering, neutral color balance, and higher resolution outputs. This is consistent with OpenAI running an A/B test of the new model against the current GPT-Image-1.5 in production, a standard practice before a full rollout.

Additionally, developers examining recent ChatGPT mobile app updates have found string references to "GPT-Image-2" in the application code, providing further evidence that a formal release is being prepared.

7 Major Capability Upgrades in GPT-Image-2

Based on the LM Arena testing data, ChatGPT A/B test reports, and community analysis, here are the most significant improvements GPT-Image-2 appears to bring over its predecessors.

1. Text Rendering Accuracy Over 99%

This is the headline improvement and the one that matters most for practical use cases.

Text rendering has been the Achilles' heel of AI image generation since its inception. Ask DALL-E 3 to put "Grand Opening" on a storefront sign and you'd get "Grnad Opennig" or something equally mangled. GPT-Image-1 improved this but still struggled with longer strings. GPT-Image-1.5 pushed accuracy to roughly 90-95%, good enough for simple labels but unreliable for anything complex.

GPT-Image-2 appears to have essentially solved this problem. In LM Arena tests, the model correctly rendered:

• Complete UI interfaces with properly spelled button text, menu items, and form labels
• Watch faces displaying specific requested times with correct hour and minute hand positions
• Multi-line text blocks with consistent fonts and proper alignment
• Product packaging with brand names, ingredient lists, and fine print

If this accuracy holds in production, it fundamentally changes what AI image generation can be used for. Social media graphics, ad creatives, presentation slides, mockups, and product images with text become viable outputs rather than exercises in frustration.

2. Yellow Color Cast Eliminated

Every version of OpenAI's image generation has exhibited a characteristic warm yellow/orange tint. It's subtle in some outputs and obvious in others, but it's been a consistent presence. Designers who use these tools regularly have developed workarounds: specifying "cool blue-toned lighting" or manually color-correcting outputs in post-production.

GPT-Image-2 outputs from LM Arena show neutral, accurate color rendering. Whites appear white. Blues appear blue. Skin tones render naturally without the warm shift. This suggests a significant change in the model's training data, color space handling, or post-processing pipeline.

For professional use cases, accurate color rendering is table stakes. This fix alone makes GPT-Image-2 substantially more useful for brand assets, product photography, and any context where color accuracy matters.

3. World Knowledge Dramatically Improved

One of the most revealing tests conducted during the LM Arena window was a Minecraft-Manhattan scene: a prompt asking the model to render a specific real-world location (Manhattan) in the visual style of another recognizable context (Minecraft). This test requires the model to simultaneously understand what Manhattan looks like, what Minecraft's visual style entails, and how to combine them coherently.

In this test, maskingtape-alpha outperformed both of its sibling models and Nano Banana Pro. The result showed recognizable Manhattan landmarks rendered in accurate Minecraft block aesthetics, with correct proportions and spatial relationships.

This improvement in world knowledge extends beyond creative mashups. It means the model has a better understanding of real-world objects, architectural styles, brand aesthetics, cultural contexts, and the relationships between them. Prompts that reference specific places, products, or styles should produce more accurate and contextually appropriate results.

4. Resolution Up to 4K Level

GPT-Image-1.5 maxes out at 1024x1024 pixels, with some upscaling options available. GPT-Image-2 is expected to support native output resolutions of at least 2048x2048, with some reports suggesting 4K capability.

Equally important is the addition of 16:9 widescreen support. This aspect ratio is essential for practical use cases that GPT-Image-1.5 handles poorly: YouTube thumbnails, presentation slides, website hero banners, LinkedIn post images, and any context designed for modern widescreen displays.

Higher resolution combined with flexible aspect ratios means fewer compromises and less post-processing. A single generation can produce a usable asset rather than a starting point that needs to be upscaled, cropped, or resized.

5. New Independent Architecture

This is perhaps the most technically significant detail to emerge. GPT-Image-2 does not appear to be built on top of GPT-4o, the multimodal model that currently handles image generation in ChatGPT. Instead, it appears to be an entirely new architecture purpose-built for image generation.

The practical implication is speed. GPT-Image-1.5, which runs through GPT-4o, often takes 10-30 seconds to generate an image depending on complexity and server load. GPT-Image-2 is expected to generate high-quality images in under 3 seconds, a dramatic improvement that would make the tool feel much more responsive and practical for iterative workflows.

A dedicated architecture also suggests that OpenAI has invested significantly in image generation as a standalone capability rather than treating it as a feature bolted onto their language model. This is a strategic signal about where they see the market heading.

6. CJK Text Rendering

One of the more surprising findings from the LM Arena tests: Chinese, Japanese, and Korean character rendering quality was described by testers as "surprisingly good." Previous OpenAI models have struggled significantly with CJK characters, often producing malformed glyphs, incorrect stroke orders, or characters that look vaguely correct but are actually nonsensical.

The GPT-Image-2 outputs showed clear, properly formed CJK characters with accurate stroke structures. If this holds up at scale, it opens the door for practical use cases in East Asian markets, including signage, packaging, social media graphics, and marketing materials in Chinese, Japanese, and Korean.

Given that CJK text rendering is substantially more complex than Latin text rendering (thousands of unique characters, precise stroke requirements, multiple writing systems), this improvement likely reflects a deliberate training effort rather than a side effect of general model improvement.

7. Multilingual Support and Complex Prompt Following

Beyond text rendering in images, GPT-Image-2 appears to handle complex, multi-part prompts with significantly greater fidelity. Prompts specifying multiple subjects with specific spatial placements, distinct colors for each element, and detailed scene compositions produced results that more faithfully matched the descriptions.

This improvement in prompt adherence applies across languages. Non-English prompts in testing showed similar levels of accuracy to English prompts, suggesting the model has been trained to understand and execute image generation instructions in multiple languages rather than routing everything through English translation first.

For global users and multilingual marketing teams, this means fewer iterations and less prompt engineering to get the desired output, a meaningful quality-of-life improvement.

Prompt adherence also matters for consistency. When running campaigns that require multiple images with a unified visual style, colors, and layout logic, a model that follows complex instructions more faithfully produces more consistent results across a batch. This reduces the number of regenerations needed and makes AI image tools more viable for production-grade visual asset pipelines.

GPT-Image-2 vs Nano Banana Pro: Head-to-Head

The LM Arena blind testing format is particularly useful because it strips away brand loyalty and expectations. Users judged outputs purely on quality. Here's how GPT-Image-2 (across its three codename variants) compared to Google's Nano Banana Pro, currently considered the leading AI image generation model.

Text Rendering

Winner: GPT-Image-2

In direct comparison, GPT-Image-2 demonstrated superior text rendering accuracy. The most cited example: a prompt requesting a watch face displaying a specific time. packingtape-alpha rendered the time correctly with accurate hand positions. Nano Banana Pro produced a watch with hands pointing to the wrong time. For any use case involving text in images, whether UI mockups, social media graphics, or product labels, GPT-Image-2 appears to have a clear edge.

Color Accuracy

Winner: GPT-Image-2

Nano Banana Pro already has good color neutrality; it doesn't suffer from the yellow cast that plagued OpenAI's models. But GPT-Image-2's elimination of its color cast means it now matches or slightly exceeds Nano Banana Pro on color accuracy. Both models produce neutral, true-to-prompt colors, but GPT-Image-2's improvement represents a bigger leap given where it started.

World Knowledge

Winner: GPT-Image-2

The Minecraft-Manhattan test was the clearest demonstration. maskingtape-alpha produced a more accurate and coherent mashup than Nano Banana Pro, correctly identifying and rendering specific Manhattan landmarks in Minecraft-style block graphics. This category tests the model's understanding of the real world, cultural references, brand aesthetics, and visual styles, an increasingly important capability as prompts become more sophisticated.

Spatial Reasoning

Winner: Nano Banana Pro

Not everything went GPT-Image-2's way. The Rubik's Cube reflection test, a prompt requesting a Rubik's Cube with an accurate mirror reflection, remains a challenge. GPT-Image-2 failed to correctly render the reflected face of the cube, getting the color arrangement wrong in the mirror. Nano Banana Pro handled this test better, suggesting it has stronger spatial reasoning and understanding of physical properties like reflections.

This matters for use cases involving product photography from multiple angles, interior design visualization, or any scene with mirrors, reflective surfaces, or complex geometric relationships.

Resolution

Winner: Tie

Both models support output resolutions up to 4K level. Nano Banana Pro has offered this capability in production for several months. GPT-Image-2 appears to match it, though we won't know the full range of supported resolutions and aspect ratios until the official release.

Speed

Winner: Competitive

GPT-Image-2 is expected to generate images in under 3 seconds, which would be competitive with Nano Banana Pro's generation times. GPT-Image-1.5's 10-30 second generation times have been a significant usability pain point, so this improvement, if confirmed, addresses one of the biggest complaints about OpenAI's image tools.

Availability

Winner: Nano Banana Pro

This one is straightforward. Nano Banana Pro is available right now. You can use it today. GPT-Image-2 has not been officially released. If you need the best available AI image generation model today, Nano Banana Pro is the answer. That will likely change within weeks, but today, availability counts for a lot.

Comparison Summary Table

The takeaway: GPT-Image-2 appears to lead in the categories that matter most for practical creative work (text rendering, color accuracy, world knowledge), while Nano Banana Pro retains an edge in spatial reasoning and, crucially, is the only one you can actually use right now.

It's worth emphasizing that these results come from blind testing where users had no idea which model they were evaluating. This removes the bias that often colors model comparisons when testers know what they're looking at. The results reflect genuine perceived quality differences, not brand preferences.

Where Does Nano Banana 2 Fit In?

While the AI image community has been focused on GPT-Image-2's LM Arena appearance, Google hasn't been standing still. On February 26, 2026, Google released Nano Banana 2, a model that combines Nano Banana Pro's image quality with Gemini Flash's speed.

Nano Banana 2 represents a different strategic approach than what OpenAI appears to be doing with GPT-Image-2. Where OpenAI is building a dedicated, standalone image generation architecture, Google is integrating image generation more deeply into its broader Gemini ecosystem. Nano Banana 2 is already rolling out across Google products, from Google Docs and Slides to Google Ads and YouTube tools.

The Three-Way Race

The competition now looks like a three-way battle:

• GPT-Image-2 — Highest raw quality (based on leaked tests), best text rendering, new dedicated architecture. Not yet available.
• Nano Banana Pro — Current quality leader in production, strong all-around performance, excellent spatial reasoning. Available now.
• Nano Banana 2 — Balances quality with speed, deeply integrated into Google's product ecosystem, optimized for high-volume use cases. Rolling out now.

Each model occupies a slightly different position. Nano Banana Pro optimizes for maximum quality. Nano Banana 2 optimizes for speed and integration. GPT-Image-2, when it launches, appears to be gunning for the quality crown while also delivering competitive speed.

It's also worth watching for how these models are priced and distributed. Google's strategy of embedding Nano Banana 2 across its product suite gives it a distribution advantage that API-only access can't match. OpenAI's strategy with GPT-Image-2 likely involves deep integration into ChatGPT, which has its own massive user base. The model that wins may not be the one with the best benchmark scores, but the one that reaches the most people in the most useful contexts.

For users and developers, this three-way competition is unambiguously good news. The pace of improvement in AI image generation is accelerating, and the rivalry between OpenAI and Google is pushing both companies to ship better models faster. The best AI image generator of 2026 will be significantly better than anything available at the start of the year.

Known Limitations and Open Questions

The hype around GPT-Image-2 is warranted based on what we've seen, but it's worth being clear about the limitations and unknowns.

Spatial Reasoning Still Needs Work

The Rubik's Cube reflection test failure is notable because it reveals a category of problems that GPT-Image-2 hasn't solved. Accurately rendering reflections, shadows at correct angles, and consistent multi-view geometry remains a challenge. For use cases like product photography (where you might want a product reflected in a glossy surface) or architectural visualization (where shadow accuracy matters), this limitation is relevant.

No Public Availability

As of April 20, 2026, GPT-Image-2 is not available to the public. The LM Arena test was brief and access was pulled quickly. The ChatGPT A/B test is reaching a small, uncontrolled subset of users. There is no API access, no waitlist, and no confirmed launch date. Everything discussed in this article is based on leaked test data and indirect evidence.

No Confirmed Pricing

OpenAI has not announced pricing for GPT-Image-2. Will it be included in ChatGPT Plus subscriptions? Will it have separate API pricing tiers? Will free-tier users get access? These questions remain unanswered. Given that the model appears to use a new, dedicated architecture rather than running through GPT-4o, the cost structure could be different from the current image generation pricing.

DALL-E 2/3 Retirement Creates Pressure

OpenAI has announced that DALL-E 2 and DALL-E 3 will be retired on May 12, 2026. This creates an interesting dynamic. Developers and applications currently using the DALL-E API will need a migration path. If GPT-Image-2 isn't ready in time, GPT-Image-1.5 (via the GPT-4o model) becomes the only option, and it's not a like-for-like replacement for all DALL-E use cases.

The retirement deadline suggests OpenAI is confident that a replacement will be available, but it also creates pressure to launch before the model may be fully polished. Whether that results in a phased rollout, a limited preview, or a full launch remains to be seen.

Safety and Content Policy Unknowns

OpenAI has historically implemented strict content policies on its image generation models. DALL-E 3 was notably conservative in what it would and wouldn't generate, frustrating many users who wanted to create legitimate content that triggered safety filters. How GPT-Image-2 handles content moderation, whether it's more or less permissive, and what its refusal patterns look like are all unknowns that will affect its practical usefulness.

Limited Real-World Testing Data

The LM Arena data comes from a window of just a few hours. The ChatGPT A/B test reports are anecdotal. We don't yet know how GPT-Image-2 performs across the full range of real-world prompts: edge cases, adversarial inputs, specific industry use cases, batch generation at scale, or consistency across multiple generations of the same prompt. Early test data is encouraging but not comprehensive.

It's also worth noting that LM Arena testing tends to favor visually impressive, creative prompts over mundane production workloads. How the model handles repetitive brand-consistency tasks, batch generation of product variants, or highly specific technical illustrations remains to be seen.

When Will GPT-Image-2 Launch?

No official launch date has been announced. But we can make an informed estimate based on the available evidence.

Historical Pattern

OpenAI has a relatively consistent pattern for major model releases. Models typically appear on testing platforms like LM Arena 2-4 weeks before public release. This pattern held for GPT-4o, GPT-Image-1, and several other recent releases. If the pattern holds for GPT-Image-2, the April 4 LM Arena appearance would put the launch window at late April to early May 2026.

The DALL-E Deadline

DALL-E 2 and DALL-E 3 are retiring on May 12. OpenAI would not retire these models without a replacement ready, especially given the number of API developers who depend on them. This strongly suggests GPT-Image-2 will be available, at least via API, by mid-May at the latest.

Mobile App Evidence

The discovery of GPT-Image-2 string references in ChatGPT's mobile app code is significant. Mobile app updates go through review processes at Apple and Google that typically take several days. Adding UI strings for a feature that's weeks or months away is unusual. This suggests the ChatGPT client-side code is being prepared for an imminent rollout.

A/B Testing in ChatGPT

The fact that the model is already being A/B tested in ChatGPT production is a strong signal. A/B testing is typically one of the final steps before a full launch. Companies use it to validate performance, catch issues, and measure user satisfaction before committing to a full rollout.

Most Likely Timeline

Taking all of this together, the most likely launch window for GPT-Image-2 is late April to mid-May 2026. A phased rollout is probable: ChatGPT Plus subscribers first, followed by API access, then broader availability. The DALL-E retirement on May 12 creates a hard deadline for API availability, even if the consumer ChatGPT rollout follows a different schedule.

There is also the possibility that OpenAI announces GPT-Image-2 alongside other product updates. The company has adopted a more frequent release cadence in 2026, with monthly announcements becoming the norm. A late April announcement event with a same-day or same-week rollout would fit both the technical evidence and OpenAI's current go-to-market strategy.

Whatever the exact date, the combination of DALL-E retirement pressure, active A/B testing, and mobile app preparation makes it clear: GPT-Image-2 is not a distant roadmap item. It's an imminent launch.

What This Means for Creators and Marketers

The competitive landscape between GPT-Image-2, Nano Banana Pro, and Nano Banana 2 is about to produce a wave of capability improvements that directly affect anyone creating visual content.

Text in Images Becomes Reliable

This is the single biggest practical change. When text rendering works consistently above 99% accuracy, entire categories of use cases open up:

• Social media graphics — Headlines, quotes, calls-to-action, and branded text overlays can be generated directly rather than added in post-production.
• Ad creatives — Banner ads, social ads, and display ads with text become one-step generations instead of multi-tool workflows.
• Product mockups — Packaging designs, label concepts, and merchandise mockups with accurate brand text can be generated in seconds for client presentations.
• Presentation slides — Illustrations with embedded text labels, charts with accurate axis labels, and diagrams with callouts become viable AI-generated assets.
• Thumbnails — YouTube thumbnails, blog post hero images, and podcast cover art with readable text can be generated without a separate design tool.

For years, the advice for AI image generation has been "generate the image, then add text in Canva/Figma/Photoshop." If GPT-Image-2 delivers on its promise, that extra step disappears for many use cases.

This shift is particularly significant for solo creators and small teams who don't have a designer on staff. The ability to generate a complete, text-included graphic in a single step removes one of the biggest friction points in content creation workflows.

Color Accuracy Opens Professional Use Cases

Eliminating the yellow color cast isn't just an aesthetic improvement. It makes AI-generated images viable for contexts where color accuracy matters: brand assets that need to match specific Pantone colors, product photography where the item's actual color matters, and marketing materials where visual consistency across channels is important.

Speed Enables Iteration

If GPT-Image-2 delivers sub-3-second generation times, the workflow changes from "generate and wait" to "generate, review, adjust, regenerate" in rapid cycles. This makes AI image generation feel more like working with a responsive design tool and less like submitting a job to a queue.

Speed matters more than most benchmarks suggest. In practice, the difference between a 3-second generation and a 20-second generation isn't just 17 seconds of wall-clock time. It's the difference between staying in a creative flow state and losing your train of thought. Faster generation means more experimentation, more variations explored, and ultimately better final outputs.

Resolution and Aspect Ratio Reduce Post-Processing

Native 4K output and 16:9 widescreen support mean that many assets can be used directly from the generator without resizing, upscaling, or cropping. A YouTube thumbnail, a blog hero image, a LinkedIn banner, or a presentation slide background can be generated at the exact dimensions needed. This eliminates an entire step from the creation workflow and reduces the risk of quality loss from post-generation resizing.

The Multi-Model Future

With GPT-Image-2, Nano Banana Pro, and Nano Banana 2 all delivering strong but differentiated capabilities, the smartest approach for serious creators is access to multiple models. Different prompts and use cases play to different models' strengths. A text-heavy social media graphic might be best served by GPT-Image-2's text rendering. A product photo with complex reflections might benefit from Nano Banana Pro's spatial reasoning. A high-volume content pipeline might optimize for Nano Banana 2's speed.

At Genra, we're closely tracking GPT-Image-2's development and plan to integrate it into our multi-model pipeline as soon as it becomes available via API. Our goal is to ensure Genra users automatically get access to the best image generation capabilities without needing to switch tools or manage multiple subscriptions. When GPT-Image-2 launches, Genra users will have it alongside Nano Banana Pro and other leading models, with intelligent routing to the best model for each specific task.

Key Takeaways

• GPT-Image-2 is OpenAI's next-generation image model. It was discovered through a brief LM Arena appearance on April 4, 2026, under the codenames maskingtape-alpha, gaffertape-alpha, and packingtape-alpha.
• The model's most significant improvement is text rendering accuracy above 99%, a quantum leap from GPT-Image-1.5's ~90-95% and a capability that opens up practical use cases like social media graphics, ad creatives, and product mockups with embedded text.
• The yellow color cast that has plagued OpenAI's image models since DALL-E is eliminated in GPT-Image-2. Color rendering is now neutral and accurate.
• In head-to-head blind tests, GPT-Image-2 beat Nano Banana Pro in text rendering, color accuracy, and world knowledge. Nano Banana Pro retained an edge in spatial reasoning.
• GPT-Image-2 uses a new, dedicated architecture (not GPT-4o), enabling sub-3-second generation times at up to 4K resolution with widescreen aspect ratio support.
• The most likely launch window is late April to mid-May 2026, driven by the DALL-E 2/3 retirement deadline on May 12 and OpenAI's historical testing-to-release timeline.
• The three-way competition between GPT-Image-2, Nano Banana Pro, and Nano Banana 2 will define the AI image generation landscape for the rest of 2026.

Frequently Asked Questions

Is GPT-Image-2 available to use right now?

No. As of April 20, 2026, GPT-Image-2 has not been officially released. It briefly appeared on LM Arena on April 4 and is currently being A/B tested within ChatGPT for a small subset of users, but there is no public access or API availability. The most likely launch window is late April to mid-May 2026.

When will GPT-Image-2 launch?

No official date has been announced. Based on OpenAI's historical pattern of 2-4 weeks from LM Arena testing to release, the DALL-E 2/3 retirement deadline on May 12, and the discovery of mobile app strings, the most likely window is late April to mid-May 2026. A phased rollout starting with ChatGPT Plus subscribers is probable.

How does GPT-Image-2 compare to Nano Banana Pro?

In blind LM Arena tests, GPT-Image-2 beat Nano Banana Pro in text rendering accuracy, color neutrality, and world knowledge. Nano Banana Pro won in spatial reasoning (the Rubik's Cube reflection test). Both support up to 4K resolution and competitive generation speeds. The key difference today: Nano Banana Pro is available now, while GPT-Image-2 is not yet released.

Will GPT-Image-2 be free?

Pricing has not been confirmed. Based on OpenAI's current model, GPT-Image-2 will likely be available to ChatGPT Plus, Team, and Enterprise subscribers with usage limits, and accessible via API with per-image pricing. Whether free-tier ChatGPT users will get access is unknown. Given the new dedicated architecture, API pricing may differ from current GPT-Image-1.5 rates.

What happened to DALL-E? Is it being replaced?

Yes. OpenAI has announced that DALL-E 2 and DALL-E 3 will be retired on May 12, 2026. GPT-Image-1 and 1.5 (integrated into GPT-4o) have already been serving as the primary image generation models in ChatGPT. GPT-Image-2 is expected to become the flagship image generation model going forward, with a new dedicated architecture rather than running through GPT-4o.

What is LM Arena and how reliable is the testing data?

LM Arena is a community-driven platform where AI models compete in blind head-to-head comparisons. Users submit prompts to two anonymous models and vote on which output they prefer. Because testers don't know which model they're evaluating, the results are considered relatively unbiased. However, the GPT-Image-2 data comes from a limited window of just a few hours, so it should be treated as promising early evidence rather than comprehensive benchmarking.

Can GPT-Image-2 render text in Chinese, Japanese, and Korean?

Based on LM Arena tests, GPT-Image-2 shows significantly improved CJK text rendering compared to previous OpenAI models. Testers described the quality as "surprisingly good" with accurate glyph forms and clear strokes. However, comprehensive testing across the full range of CJK characters and font styles hasn't been possible given the limited test window.

What is Nano Banana 2 and how does it differ from Nano Banana Pro?

Nano Banana 2 is Google's latest image generation model, released on February 26, 2026. It combines Nano Banana Pro's image quality with Gemini Flash's speed, optimizing for fast generation times and deep integration across Google products. Think of Nano Banana Pro as the quality-focused model and Nano Banana 2 as the speed-and-integration-focused model. Both are available now.