Nano Banana had pushed Gemini to the mainstream last year, and three similarly cryptically-named models are now creating buzz on Arena — and they are rumoured to be from the OpenAI stable.
The three models — maskingtape-alpha, gaffertape-alpha, and packingtape-alpha — appeared on Arena with little fanfare, but the AI community quickly took notice. Developer and indie hacker Pieter Levels (@levelsio) was among the first to flag them, posting that the models appeared to have “extremely good world knowledge and great text rendering.” Venture investor Justine Moore (@venturetwins) echoed the sentiment, highlighting two prompts in particular — “average engineer’s screen” and “young woman taking selfie with Sam Altman” — as evidence of unusually strong world knowledge. The Sam Altman detail, if accurate, is a telling clue about provenance.
Community testing has surfaced more specifics. One user noted that packingtape-alpha correctly rendered the time on a watch in an image, something Nano Banana Pro failed to do. Another found that in a side-by-side comparison of a first-person Minecraft gameplay screenshot set in Manhattan, maskingtape-alpha outperformed all three of its tape siblings — and Nano Banana Pro. Not everything is perfect, though: the models reportedly still fail the Rubik’s Cube reflection test, a standard stress test for spatial reasoning in image generation.
The Nano Banana Playbook
The tape trio’s Arena debut follows a now-familiar script. Nano Banana first appeared on LMArena in August 2025 with no company claiming credit, quickly drawing attention for its image editing capabilities — particularly its ability to preserve character consistency and follow complex instructions. Within days, it had shot to the top of the leaderboard. Google hinted at its involvement through a series of banana emoji posts by senior executives including Demis Hassabis, before formally confirming the model was Gemini 2.5 Flash Image.
What followed was a genuine cultural moment. Nano Banana brought 10 million new users to Gemini and facilitated over 200 million image edits within weeks. Gemini briefly became the top free app on the App Store, displacing ChatGPT. The model’s virality was driven in large part by the figurine trend — users generating 3D-style miniature figures of themselves — which spread particularly fast in India and Southeast Asia.
Crucially, Google leaned into the codename rather than retiring it. What started as a 2:30 a.m. improvisation by product manager Naina Raisinghani became the brand. The Nano Banana name has since persisted through two model generations — Nano Banana Pro (Gemini 3 Pro Image) and Nano Banana 2 (Gemini 3.1 Flash Image) — with Google explicitly choosing continuity over conventional naming.
OpenAI’s Image Ambitions After Sora
The timing of the tape models’ appearance is notable. OpenAI announced the discontinuation of Sora on March 24, 2026, just six months after launching it as a standalone app. The shutdown was abrupt enough that Disney — which had committed to a $1 billion partnership built around the platform — found out less than an hour before the public announcement. The official reason given was reallocation of compute toward “world simulation research to advance robotics,” but the subtext was clear: Sora was burning roughly $1 million per day with a user base that had collapsed from a peak of around one million to under 500,000.
Killing Sora freed up significant compute. The question is where that capacity goes. A next-generation image model is one obvious answer. OpenAI’s GPT Image 1.5 had already topped the LMArena image leaderboard in December 2025, edging past Nano Banana Pro — a direct response to Google’s momentum in the space. A GPT-Image-2, if that’s what the tape models represent, would be OpenAI doubling down on the one consumer AI category where viral, mainstream adoption remains genuinely achievable.
The naming convention is consistent with OpenAI’s existing gpt-image family. And the decision to test under adhesive-tape pseudonyms on Arena — rather than through a controlled API rollout — mirrors exactly how Google stress-tested Nano Banana before its public launch.
What to Watch
Arena’s blind testing methodology is the point. Models are judged by users who don’t know which company produced the output. If maskingtape-alpha and its siblings hold up under sustained community testing, the Elo scores will make the case without any marketing spend. That’s precisely what happened with Nano Banana — it amassed over 2.5 million votes and built the largest Elo score lead in Arena history before Google said a word publicly.
OpenAI hasn’t confirmed anything. But if the tape models do turn out to be GPT-Image-2, the company will have successfully used the same Arena-first playbook that caught it flat-footed when Google deployed it eight months ago.