GPT-IMAGE-2 Likely on LMarena
A mysterious trio of models named after tape types outperformed competitors and claimed OpenAI origin.
A significant leak on the popular AI benchmarking platform LMSys Chatbot Arena has potentially revealed OpenAI's next-generation multimodal model, tentatively dubbed GPT-Image-2. The model appeared under three distinct codenames—maskingtape-alpha, gaffertape-alpha, and packingtape-alpha—suggesting different configurations or sizes. Early testers who managed to interact with the models before their swift removal reported exceptional performance, with one user stating it was 'far better than nano banana,' a reference to a previous high-performing model. Crucially, when prompted, the AI reportedly identified itself as an OpenAI product, strongly indicating its origin.
The rapid removal of all three models from the public arena by LMSys administrators is a classic telltale sign of an accidental early release, often preceding an official announcement. This pattern mirrors previous leaks of major AI models. The incident has ignited speculation within the AI community that OpenAI is preparing to launch a successor to its GPT-4V (Vision) model, potentially with significantly enhanced image understanding, reasoning, and generation capabilities. The use of tape-related codenames follows OpenAI's quirky naming tradition for internal projects.
- Three models (maskingtape/gaffertape/packingtape-alpha) leaked on LMSys Arena, claiming to be from OpenAI.
- Early testers reported performance 'far better' than the previous 'nano banana' benchmark standard.
- All models were swiftly removed, a strong indicator of an accidental leak ahead of a planned launch.
Why It Matters
A leak of this magnitude suggests a major leap in multimodal AI is imminent, potentially reshaping competitive benchmarks.