GPT Image 2: What leaked tests and reports claim (April 2026)

Status: OpenAI has not shipped a product officially named “GPT Image 2” or published a spec sheet for it. Everything below is third-party reporting and community inference—useful context, not confirmation. We link sources so you can judge them yourself.

#What surfaced

In early April 2026, several independent outlets and creators reported that OpenAI was quietly testing a next-generation image model—often referred to in commentary as GPT Image 2, GPT-Image-2, or internally as “Image V2”—via public blind-test venues and ChatGPT.

The story that gained traction runs like this: three image-generation entries appeared on LMSYS / LM Arena under hardware-store-style codenames, performed strongly in head-to-head comparisons, then disappeared from the public leaderboard after people connected the dots. Separately, some ChatGPT users reported seeing new image outputs, A/B comparisons between results, or broader access—also described as unconfirmed and uneven.

We are not asserting these accounts are complete or accurate; we are summarizing what has been written publicly.

#The Arena codenames (community-labeled)

Reports consistently name three test labels (exact spelling varies by screenshot):

• maskingtape-alpha
• gaffertape-alpha
• packingtape-alpha

Testing Catalog states OpenAI appears to be testing what it refers to internally as Image V2 in three variants under those names, notes that images were pulled back from the Arena but could still appear in ChatGPT, and draws a parallel to December 2025 Arena previews under codenames Chestnut and Hazelnut—which, per that outlet, shipped shortly afterward as GPT Image 1.5. That historical parallel is reason some observers expect a near-term official drop; it is not a guarantee.

Other write-ups (Office Chai, Oimi AI’s summary, MagicShot) repeat similar details with varying emphasis—always treat them as journalism about rumors, not primary sources.

#What testers and bloggers claim (capabilities)

Across those reports, the claimed improvements—not verified by OpenAI here—include:

• Stronger text-in-image — readable UI copy, fewer “gibberish button” failures; some pieces explicitly compare the narrative to Google’s Nano Banana Pro tier.
• Photorealism and lighting — more convincing materials, reflections, and scenes in anecdotal samples.
• “World knowledge” in the picture — e.g. plausible brand-like UIs, clocks, maps, or product context; still subject to classic generator failure modes (some articles mention tests like Rubik’s cube reflections as still problematic in samples).

Again: these are impressions from leaks and social buzz, not a Kubeez benchmark.

#Official silence and release timing

As of this writing, OpenAI has not announced GPT Image 2, an API model id for it, or a release date. Any ChatGPT rollout, A/B test, or API plan remains speculation until the company posts pricing, policy, and model cards.

#Why this matters for teams

If the reports are directionally right, the next OpenAI image tier could raise the bar for marketing layouts, UI mocks, and text-heavy creatives—the same battleground where Nano Banana Pro and others have been winning leaderboard attention. Pricing, safety filters, and latency at launch could still differ from what people saw in previews (Testing Catalog discusses that possibility explicitly).

#Kubeez

We add new models when they are stable, licensed, and wired into our stack—watch the in-app model list and our AI models guide. For production image work today, use Images and Media generate with the currently available tiers shown in the product.

Bottom line: GPT Image 2 is a label the commentary ecosystem is using for unreleased OpenAI image tech spotted in Arena-style testing and ChatGPT rumors in April 2026. Treat every claim as unverified until OpenAI publishes the real name, access rules, and terms.

#Further reading (external)

• Testing Catalog — OpenAI tests next-gen Image V2 on ChatGPT and LM Arena
• Office Chai — Maskingtape / Gaffertape / Packingtape on Arena