OpenAI’s next image generation model has been spotted in the wild. On April 4, 2026, three anonymous image models appeared on LM Arena under bizarre codenames, generated images that left testers stunned, and then vanished within hours.
Everything points to GPT Image 2.
The leaked models showed near-perfect text rendering, eliminated the yellow color cast that plagued GPT Image 1.5, and beat Google’s Nano Banana Pro in blind comparisons. With DALL-E shutting down on May 12, 2026, OpenAI clearly has a replacement ready to go.
How OpenAI’s Image Models Evolved
OpenAI’s image generation journey started with DALL-E back in 2021, but the real shift happened in March 2025 when the company released GPT Image 1. Unlike DALL-E, which was a separate diffusion-based model called through a tool, GPT Image 1 generated images natively inside the language model itself. It was autoregressive, meaning it produced images token by token, just like text.
The impact was immediate. Over 130 million users created more than 700 million images in the first week alone.
Then came GPT Image 1 Mini in October 2025 at OpenAI DevDay, offering 80% cheaper API pricing. Two months later, in December 2025, OpenAI launched GPT Image 1.5 with up to 4x faster generation and 20% cheaper API costs. That model climbed to the top of the LM Arena leaderboard with an Elo rating of 1264.
Here’s the full timeline:
| Model | Release Date | Key Improvement |
|---|---|---|
| DALL-E 2 | April 2022 | First widely available AI image generator |
| DALL-E 3 | October 2023 | Better prompt following, ChatGPT integration |
| GPT Image 1 | March 25, 2025 | First autoregressive image model inside ChatGPT |
| GPT Image 1 Mini | October 6, 2025 | 80% cheaper API pricing |
| GPT Image 1.5 | December 16, 2025 | 4x faster, better quality, new Arena #1 |
| GPT Image 2 | Coming soon | New architecture, near-perfect text, no color cast |
The LM Arena Leak
On April 4, 2026, three anonymous image models showed up on LM Arena (the platform formerly known as Chatbot Arena where users compare AI models in blind tests). They appeared under the codenames:
- maskingtape-alpha
- gaffertape-alpha
- packingtape-alpha
Developer Pieter Levels and venture investor Justine Moore were among the first to publicly flag the models. Several other community members, including @Angaisb_, @synthwavedd, and @blakeir, tested them extensively before they disappeared.
All three models were pulled from the Arena within hours.
This pattern is not new. In December 2025, two anonymous models appeared on LM Arena under the codenames “Chestnut” and “Hazelnut.” They were tested briefly, removed, and weeks later OpenAI shipped GPT Image 1.5. The tape models follow the exact same playbook.
Beyond the Arena testing, some ChatGPT users reported being randomly served a noticeably different (and better) image model during regular usage, suggesting active A/B testing in the product itself.
What the Leaked Models Can Do
Testers who caught the tape models before they vanished documented several capabilities that represent a significant leap over GPT Image 1.5.
Text Rendering Is Nearly Perfect
This has been the biggest weakness of every AI image generator. GPT Image 1.5 achieved roughly 90-95% accuracy on text in images, which sounds good until you see misspelled signs and garbled labels in your outputs.
The leaked models produced text that testers described as “near-perfect” and “finally usable.” Text appeared naturally on signs, product labels, code snippets, UI mockups, and even handwritten medical notes with convincing penmanship. Comic book panels had readable speech bubbles. Watch faces showed correctly positioned hands matching the time described in the prompt.
The Yellow Color Cast Is Gone
GPT Image 1.5 had a well-documented warm yellow tint that affected most outputs. Photographers and designers consistently complained about it. The tape models produced images with natural, neutral color rendering. No post-processing needed.
Photorealism That Fools You
Portraits generated by the leaked models were described as “indistinguishable from real photographs.” Complex beach scenes showed accurate hand anatomy (historically a weak point for AI image generators) and realistic sunglass reflections. This is a noticeable jump from GPT Image 1.5, which still produced occasional uncanny-valley artifacts.
World Knowledge and Detail
The models demonstrated an understanding of the real world that went well beyond prompt comprehension. Testers generated IKEA storefronts with architectural accuracy, YouTube and Windows interfaces realistic enough to pass as screenshots, Minecraft scenes with correct in-game UI and art style, and geographically accurate world maps with proper labels and topographic shading.
Wider Aspect Ratios
GPT Image 1.5 was limited to three aspect ratios: 1:1, 3:2, and 2:3. The leaked models confirmed support for 16:9, which makes them far more useful for presentations, thumbnails, and video content.
What Still Doesn’t Work
The models are not flawless. A Rubik’s Cube mirror reflection test stumped them, which is an industry-wide challenge that no current image model handles correctly. But on virtually every other benchmark, the tape models outperformed everything currently available.
GPT Image 2 vs GPT Image 1.5
Based on what testers observed from the leaked models, here’s how GPT Image 2 stacks up against its predecessor:
| Feature | GPT Image 1.5 | GPT Image 2 (Leaked) |
|---|---|---|
| Text accuracy | ~90-95% | ~99%+ |
| Aspect ratios | 1:1, 3:2, 2:3 | Adds 16:9, likely 9:16 |
| Color rendering | Warm yellow tint | Natural, neutral colors |
| Max resolution | 1536×1024 | 2048×2048+ (expected) |
| Speed | 8-12 seconds | Under 3 seconds (expected) |
| Architecture | Based on GPT-4o pipeline | Brand new standalone model |
| Non-Latin scripts | Poor | Significantly improved |
| Character consistency | Drift across generations | Likely persistent embeddings |
| Photorealism | Good | Near-photographic |
| Hand anatomy | Occasional errors | Highly accurate |
The jump in text accuracy alone makes this a meaningful upgrade for anyone using AI image generation for marketing materials, social media graphics, or product mockups.
A Completely New Architecture
One of the most interesting technical details from the leak is that GPT Image 2 appears to be built on an entirely new architecture. Previous GPT Image models were built on top of GPT-4o’s image pipeline. GPT Image 2 is reportedly a standalone model.
Multiple sources report a shift from a two-stage inference process to single-pass inference, which would explain both the quality improvement and the expected speed gains. New metadata tags have also been detected in PNG file outputs from suspected GPT Image 2 generations, further supporting the theory of a fundamentally different system.
This architectural overhaul matters because it suggests OpenAI is not just iterating. They’re rebuilding the image generation stack from the ground up.
How It Stacks Up Against Competitors
The AI image generation space has gotten extremely competitive. Here’s where GPT Image 2 appears to land based on the leaked performance:
| Model | Strength | GPT Image 2 Comparison |
|---|---|---|
| Midjourney V7/V8 | Artistic quality, aesthetic control | GPT Image 2 leads on text and prompt-following. Midjourney still stronger on artistic style. |
| Google Nano Banana Pro | Previous LM Arena leader | Tape models beat it in blind tests on realism, text, and world knowledge. |
| Nano Banana 2 | Google’s latest, 4K capable | Close competitor, but GPT Image 2 appears stronger on text rendering. |
| FLUX.2 | Open-source, customizable | More flexible for developers, but GPT Image 2 integrates better with chat. |
| Adobe Firefly | Enterprise, brand safety | Purpose-built for commercial use. GPT Image 2 more generalist. |
| Seedream 4.5 | Chinese model, strong quality | Competitive on quality, but lacks ChatGPT ecosystem integration. |
The most telling result: in blind Arena comparisons, testers noted that the tape models made Nano Banana Pro “look like DALL-E.” One tester said they “outperform NBP in realism, text rendering, and world knowledge simultaneously.” That’s a strong statement given that Nano Banana Pro had dominated the image generation leaderboard for months.
DALL-E Is Officially Dead
OpenAI announced in November 2025 that both DALL-E 2 and DALL-E 3 will be shut down on May 12, 2026. Azure OpenAI already retired DALL-E 3 on February 18, 2026.
The official replacement is the GPT Image model family. OpenAI specifically listed GPT Image 1 Mini as the DALL-E 3 replacement for API users. But with GPT Image 2 clearly in late-stage testing, it seems likely that the real successor will arrive right around the DALL-E shutdown date.
This creates an interesting timing pressure. Developers currently using DALL-E 3 through the API need to migrate by May 12. If GPT Image 2 launches before that deadline, it gives them a clear upgrade path rather than a sideways move to GPT Image 1 Mini.
Sora’s Shutdown Freed Massive Compute
There’s a backstory here that helps explain the timing. OpenAI shut down Sora, its AI video generation tool, on March 24, 2026. The numbers were brutal: Sora was burning $15 million per day in inference costs at peak while generating only $2.1 million in total lifetime revenue.
Sam Altman stated the shutdown was to “focus compute and product capabilities on next generation” products. That freed up an enormous amount of GPU capacity, and it’s hard to imagine that capacity isn’t being redirected toward models like GPT Image 2.
The timing lines up perfectly. Sora shut down March 24. The tape models appeared on LM Arena eleven days later on April 4. That’s not a coincidence.
Expected Pricing
GPT Image 1.5 currently costs $0.133 to $0.200 per image through the API, depending on resolution and quality settings. GPT Image 1 Mini runs about 80% cheaper.
Industry analysts expect GPT Image 2 to land around $0.15 to $0.20 per image, representing a slight increase from GPT Image 1.5. The new standalone architecture likely requires more compute per generation, which would justify the higher price point.
For ChatGPT subscribers, the model will almost certainly be included in Plus, Pro, and Go plans, just as current image generation is today.
| Tier | Estimated Cost |
|---|---|
| ChatGPT Free | Limited generations per day |
| ChatGPT Go / Plus | Included in subscription |
| ChatGPT Pro | Unlimited generations |
| API (standard) | ~$0.15-0.20 per image |
| API (mini variant) | TBD, likely 60-80% cheaper |
When Will GPT Image 2 Launch
Everything points to an imminent release. Here are the key signals:
The testing pattern matches GPT Image 1.5. In December 2025, anonymous models (“Chestnut” and “Hazelnut”) appeared on LM Arena, were pulled quickly, and GPT Image 1.5 launched within weeks. The tape models follow the same pattern.
A/B testing is active in ChatGPT. Some users are already seeing a different, better image model in their regular ChatGPT sessions. OpenAI doesn’t A/B test products that are months away from launch.
The DALL-E deadline creates urgency. DALL-E 2 and 3 shut down May 12, 2026. Launching GPT Image 2 before that date gives developers a clear migration target.
Sora compute is now available. The March 24 Sora shutdown freed massive GPU resources, enabling the scale needed for a new image model rollout.
Three simultaneous test variants suggest final evaluation. Testing three model variants at once (maskingtape, gaffertape, packingtape) suggests OpenAI is comparing final candidates, not early prototypes.
The most likely window is late April to mid-May 2026. Some analysts speculate it could launch alongside a GPT-5.4 update, but that’s unconfirmed.
What to Expect From GPT Image 2
Beyond what the leaks have confirmed, here are the capabilities that multiple credible sources expect:
Higher resolution output. GPT Image 1.5 maxes out at 1536×1024. Competitors like Midjourney and Nano Banana 2 already support 2K or higher. GPT Image 2 is expected to offer at least 2048×2048, possibly up to 4K.
Multilingual text rendering. GPT Image 1.5 struggles with non-Latin scripts like Chinese, Arabic, and Hebrew. Early tests of the tape models showed improved accuracy with Turkish text, and broader multilingual support is expected at launch.
Faster generation speed. The single-pass architecture should bring generation time down from 8-12 seconds to under 3 seconds for standard outputs.
Character consistency. One of the most requested features is maintaining a consistent character identity across multiple image generations. GPT Image 2 may introduce persistent embeddings or a reference system to solve this, though this remains speculative.
Region-based prompting. Some rumors point to ControlNet-style spatial controls like bounding boxes or layered composition tools. This would give users much more precise control over where elements appear in the image.
How to Try GPT Image 2
Once GPT Image 2 officially launches, the fastest way to access it will be through ChatGPT with a Plus, Pro, or Go subscription. If you want to compare it across multiple AI models in one place, Fello AI bundles ChatGPT alongside Claude, Gemini, and other models in a single app for Mac, iPhone, and iPad.
For developers, API access will likely follow the same rollout pattern as previous models: ChatGPT first, then API access 2-4 weeks later.
FAQ
Is GPT Image 2 the same as DALL-E 4?
No. OpenAI has moved away from the DALL-E branding entirely. Both DALL-E 2 and DALL-E 3 are being shut down on May 12, 2026. The GPT Image model family is architecturally different from DALL-E, using autoregressive generation inside the language model rather than a separate diffusion model.
When will GPT Image 2 be released?
Based on the LM Arena leak on April 4, 2026, active A/B testing in ChatGPT, and the DALL-E shutdown deadline of May 12, most analysts expect GPT Image 2 to launch between late April and mid-May 2026.
Will GPT Image 2 be free?
GPT Image 2 will likely be available to free ChatGPT users with daily generation limits, just like GPT Image 1.5 today. Full access will require a ChatGPT subscription (Go, Plus, or Pro). API pricing is expected to be around $0.15-0.20 per image.
What are maskingtape, gaffertape, and packingtape?
These are codenames for three anonymous image models that appeared on LM Arena on April 4, 2026. They are widely believed to be test variants of GPT Image 2. All three were removed from the platform within hours, following the same pattern OpenAI used when testing GPT Image 1.5 under the codenames “Chestnut” and “Hazelnut” in December 2025.
Is GPT Image 2 better than Midjourney?
Based on leaked Arena results, GPT Image 2 appears to outperform Midjourney on text rendering, instruction-following, photorealism, and world knowledge. Midjourney still holds an edge in artistic style and aesthetic control. For practical use cases like marketing materials, product mockups, and social media graphics, GPT Image 2 looks like the stronger choice.
Is GPT Image 2 better than Nano Banana Pro?
In blind comparisons on LM Arena, the leaked GPT Image 2 models consistently beat Nano Banana Pro on realism, text rendering, and world knowledge. One tester noted the difference “makes NBP look like DALL-E.” Google has since released Nano Banana 2, which narrows the gap, but GPT Image 2 still appears to lead on text accuracy.
Can GPT Image 2 generate text in images?
Yes, and this is one of its biggest improvements. While GPT Image 1.5 achieved roughly 90-95% text accuracy, the leaked GPT Image 2 models demonstrated near-perfect text rendering on signs, labels, UI elements, code snippets, and even handwritten text.
