OpenAI is preparing Image v2. New strings referencing the model have surfaced in its mobile app. At the same time the Image v2 models returned to testing on LM Arena. These signals line up and indicate the release sits close.
The codenames packingtape-alpha maskingtape-alpha and gaffertape-alpha first drew attention earlier this month for their handling of text inside generated scenes. Initial runs produced outputs where words sat naturally without the distortions common in GPT Image 1.5. After removal the models came back with adjustments that addressed voter feedback from the arena. This test adjust retest loop matches exactly how OpenAI has shipped other recent updates.
What the Arena Tests Show
LM Arena presents blind side-by-side comparisons. Voters rank results on prompt adherence and overall quality. Image v2 entries stood out on prompts requiring precise text placement within complex compositions. Earlier versions often warped letters or dropped them entirely from the scene. The current tests demonstrate clear gains on that specific weakness. Separate user reports confirm limited A/B test access inside ChatGPT where the model delivers outputs consistent with the arena performance.
The return after the initial pull suggests OpenAI incorporated the early data. Mobile app strings now reference Image v2 components by name. Code like this appears when backend systems connect to user-facing features across iOS and Android. It means the rollout path includes direct integration into the existing ChatGPT image generation tools rather than a new interface.
This chart aggregates scores from arena samples and community tests. Text rendering accuracy measures how often generated words match the prompt exactly in style placement and legibility. The jump to 95 percent for Image v2 reflects fewer manual corrections needed afterward.
Why Text Rendering Gains Matter for Real Work
Design teams that produce dozens of variants for campaigns or presentations lose hours fixing letter distortions and composition errors. Image v2 tightens prompt following and reduces stray artifacts while preserving OpenAI’s characteristic style. The improvement compounds. Marketing groups can generate branded assets with accurate taglines embedded in scenes without post-production fixes. This does not rewrite every workflow but it removes a persistent point of friction that has limited image generation adoption at scale.
I have reviewed the arena examples. The difference appears most clearly on prompts that combine objects with specific overlaid text. Previous models required careful prompt engineering or heavy editing. The samples suggest Image v2 handles those cases more reliably. For a longer view on how OpenAI image outputs have shifted review the comparison at AI Images 2014 vs 2026. Image v2 extends the style focus while fixing the technical gaps that mattered most to users.
How This Release Fits the Current Pattern
The signals follow the same sequence seen across recent model drops. Internal references appear first. Limited testing follows. Infrastructure updates surface in the mobile app. Then access widens through A/B groups before full availability. The mobile strings serve as one of the strongest indicators because they show the product integration work has reached final stages. Watch for expansion of the A/B test pools in the coming days. Additional arena entries under the same codename family will confirm whether the gains hold at higher volumes.
My assessment based on the visible data is that the preparation has advanced far enough for a near-term window. The focus remains on iterative capability increases rather than any fundamental rewrite of image generation. That approach has delivered consistent value in prior releases. Teams relying on visual content for client work or social campaigns stand to benefit first. The text handling upgrades alone justify monitoring the public leaderboards and app update notes closely. Further LM Arena data will reveal consistency once usage scales. Mobile users will likely encounter the feature directly in the ChatGPT image tools once the backend switches activate. The pattern holds: leaks and tests precede access by a short period. Image v2 aligns with that sequence.
Continued observation of both the arena and app code changes will clarify the exact timing. The combination of resumed testing and mobile integration references gives concrete direction. This update builds directly on the foundation of GPT Image 1.5 by targeting its clearest shortcomings. The result should translate into faster iteration cycles for anyone whose output depends on accurate text within visuals. The infrastructure sits in place. The test results align. The next phase looks close.