OpenAI is hosting a livestream today at 12pm PT, and the company's teaser post may have already given away the headline feature. The promotional image shows a ChatGPT conversation displayed on a Mac desktop, explicitly captioned "This is not a screenshot" to signal the entire scene was AI-generated rather than captured from a real screen.
The manufactured chat depicts a user asking about interesting events happening today, with ChatGPT responding that OpenAI is hosting a livestream. The meta-marketing move accomplishes two things at once: it promotes the event while demonstrating a level of photorealism and UI accuracy that OpenAI's current public image tools cannot reliably achieve.
The GPT-Image-2 Evidence
GPT-Image-2 has been spotted in A/B tests inside ChatGPT over the past several weeks. The early signals are genuinely interesting: near-perfect text rendering inside images, realistic UI screenshots, and a noticeable step up in photorealism. Real users have been served the new model without knowing it, and the output differences are visible enough that the AI community has started cataloguing them.
In early April, three anonymous image generation models suddenly appeared on the LM Arena evaluation platform, codenamed maskingtape-alpha, gaffertape-alpha, and packingtape-alpha. The stunning image quality they displayed shocked testers. A few hours later, these three models vanished from the platform. The community quickly concluded: this is OpenAI's upcoming GPT-Image-2.
The leaked model addressed several persistent weaknesses. GPT-Image-1 had real problems: yellow tint, inconsistent text, hands that still broke in complex scenes. The early GPT-Image-2 examples suggest most of those problems are solved.
What About ChatGPT 5.5?
Some observers expect today's event to include a model upgrade beyond image generation. The company's next major AI model, codenamed "Spud," likely to ship as the upcoming ChatGPT 5.5, could be released much earlier than we anticipated. The pretraining for the model was completed last month.
But the evidence for 5.5 is thinner than for the image model. As of April 12, 2026, OpenAI has published official material for GPT-4.5, GPT-5, GPT-5.1, GPT-5.3, and GPT-5.4 across ChatGPT and the API. The term exists in rumor posts, speculative blog coverage, social chatter, and community references, but the public OpenAI record points somewhere else.
The safest bet remains the image model. OpenAI's teaser post functions as its own proof of concept. Desktop elements, weather data, accurate interface chrome, legible text at multiple sizes. These are precisely the capabilities the leaked model demonstrated, and exactly what the current GPT-Image-1.5 struggles to nail consistently.
Timeline Pressure
OpenAI has officially announced that DALL-E 2/3 will be fully retired on May 12, 2026. The company needs a replacement ready before then. OpenAI has begun a staggered rollout of a new image generation model to ChatGPT Plus and Pro subscribers. While the company has not issued an official announcement, social media reports from paid users indicate the model is active as of April 19, 2026.
An omnimodal update bundling improved native image generation with reasoning improvements is also possible. But no strong signals have emerged for major new agents or browser capabilities at this particular event.
Everything remains speculation until the stream begins. Anthropic's recent Claude Design launch shows the competitive pressure in AI-native creative tools is intensifying. OpenAI's response may be sitting in that "not a screenshot" image, waiting to go live.
