OpenAI is quietly testing GPT Image 2

In the first week of April, anonymous image generation models named packingtape-alpha, maskingtape-alpha, and gaffertape-alpha appeared on LM Arena, and the community quickly identified them as OpenAI's GPT Image 2. OpenAI took them down within 48 hours, but by then their performance had already been exposed through public benchmarks.

Internally it is reportedly called Image V2, and related strings have reportedly been found in ChatGPT response headers and the mobile app binary as well. There are also signs that some ChatGPT Plus/Pro users are already receiving new outputs in the form of A/B testing.

There are four key improvements.

• Text rendering: generates clear paragraphs, UI labels, code snippets, and even CJK characters
• Improved color: reduces the yellowish tint of existing GPT Image 1/1.5 and outputs a more photorealistic color profile
• World knowledge: depicts real products, interfaces, and environments more accurately
• Speed: about 2x faster than GPT Image 1.5 based on community estimates, with under 3 seconds of latency

Behind this are the overlapping factors of DALL-E 2/3's shutdown on May 12, GPU capacity freed up by Sora's suspension, and a reshuffle of OpenAI's executive team. The competitive landscape is also fierce, with Google Nano Banana 2, Midjourney V8 Alpha, Flux 2, and Adobe Firefly Image 5 each already asserting their own strengths.

However, GPT Image 2 has not yet officially launched, so there is no API or pricing sheet, and only some users are noticing the quality improvements. Even so, in practical areas like images containing text, product mockups, UI drafts, and infographics, it is likely to become a much more useful tool than existing models.