OpenAI testing Image V2 (early sighting)
Community examples show OpenAI is testing a next‑generation Image V2 model inside ChatGPT and on LM Arena, according to early sightings. The report notes pricing will matter because past image releases have shifted unit economics — Image 1.5 was about 20% cheaper than its predecessor — and testing does not guarantee a public launch. For teams building image features, this is a reminder that model quality, safety filters and API pricing can all change between beta and general availability. (testingcatalog.com)
OpenAI appears to be quietly testing a new image model inside ChatGPT and on LM Arena, the public benchmark site where users compare anonymous model outputs side by side. The evidence is not an announcement. It is a trail of sightings. TestingCatalog reported on April 6 that the model surfaced in three variants on LM Arena, was later pulled back, and still appears for some ChatGPT users through A/B tests that ask them to choose between competing images. That is enough to show experimentation is happening. It is not enough to prove a launch is near. What makes the sighting interesting is not just that a new model exists. It is what people say it does better. Early testers pointed to two old weaknesses in AI image generation that still matter in practice: following prompts exactly, and rendering interface elements with legible text. TestingCatalog says the new model, referred to internally as Image V2, produced realistic UI mockups with correctly spelled button labels and stronger compositional control than users expected from OpenAI’s current public image stack. That is a narrow claim, but it matters because bad text rendering is one of the fastest ways an image model reveals its limits. That also explains why the model showed up on LM Arena in the first place. Arena testing is useful when a company wants fast, messy, human preference data instead of a polished product reveal. OpenAI has used this pattern before. TestingCatalog says the company ran similar blind tests in December 2025 with image models codenamed Chestnut and Hazelnut, and those experiments were later followed by the release of GPT Image 1.5. The lesson is not that every test ships. It is that OpenAI now treats public benchmarking as part of product development. The business side matters as much as the visuals. OpenAI’s current flagship image model in the API is GPT Image 1.5, and its pricing shows how small model revisions can reshape costs for developers. OpenAI lists GPT Image 1 at $0.011, $0.042, and $0.167 per 1024×1024 image at low, medium, and high quality. GPT Image 1.5 drops those same tiers to $0.009, $0.034, and $0.133. That is roughly a 20 percent cut across the board. If Image V2 improves quality but raises inference costs, teams that generate images at scale will feel it immediately. If it improves quality while holding or lowering price, it changes the market. That is why the early screenshots matter less than the deployment path. OpenAI’s own documentation says GPT Image 1.5 is the company’s latest production image model today, available through the Responses API and Image API, with snapshots that let developers lock behavior for consistency. An unreleased model inside ChatGPT does not offer any of that stability yet. Safety behavior can change. Output style can change. A model that looks unusually capable in a weekend test can be toned down before release, either to reduce cost or to tighten moderation. For developers, the real signal is simple. OpenAI is still actively tuning image generation in public view, and it is doing so in the places where users notice quality first: readable text, structured layouts, and prompt fidelity. On the API side, the current production benchmark remains GPT Image 1.5, with high-quality 1024×1024 generations priced at $0.133 each.
