RTK

Anomaly who makes T2I models sweat.

[RTK @T2IRedTeamKing]

If you want to read something interesting, here is a link about the Leaderboard Illusion (Arena AI): https://t.co/ODGveutWT3

I respect whoever or the group of people who made this. They are doing a great job exposing the distortion that has been going on for far too long in AI.

[RTK @T2IRedTeamKing]

True accountability will only come if a competing platform forces Arena AI change or if the labs themselves realize that buying biased, low-effort user preference data is causing their next-generation models to regress.

[RTK @T2IRedTeamKing]

People need a platform that is Anti-Arena—one built strictly for rigorous technical benchmarking, instruction adherence, and architectural transparency instead of flashy, low-effort popularity contests. This is what the AI industry desperately needs.

[RTK @T2IRedTeamKing]

As long as current leadership remains focused on maintaining elite gatekeeper status for corporate keynotes, Arena AI will continue to reward hyper-stylized, superficial outputs over engineering rigor and people just kissing ass rather than trying to search the truth.

[RTK @T2IRedTeamKing]

Here’s the question: What is the point of “testing” a model if the complex prompts you throw at it results in those SSW errors? As if Arena AI treats budget like they are starting out of the home of their parents. Shouldn’t a frontier not be like that if AI wants to be better? 🤔

[RTK @T2IRedTeamKing]

Tested GPT Image Sunburst 2.5 at least more than 3 times with a complex T2I prompt on Arena AI’s Direct Chat. Results in Something Went Wrong errors. Can’t be the prompt as the problem…nor is it cookies, force closing, etc. That got me curious.

[RTK @T2IRedTeamKing]

Interestingly enough, T2I models on Anonymous Battle that aren’t expensive are able to render that complex prompt. However, I sense that models that are expensive trigger SSW errors. A cheap model has no issue rendering a 5-reference T2I prompt, but an expensive one is different.

[RTK @T2IRedTeamKing]

The same can be said about a what-if scenario to bring back Meta’s muse-image on Direct Chat for 2 weeks (the usual Arena AI limited time if it is not them tweeting 48 hours), as I seen this at least more than once sporadically on Anonymous Battle.

[RTK @T2IRedTeamKing]

Don’t be surprised if or when Arena AI does the same thing regarding the what-if Nano Banana 3 Pro (assuming Google releases this between Nov-Dec 2026).

My advice to people: Treat models on Direct Chat like it is the last since you never know when they get pulled unexpectedly.

[Arena.ai @arena]

For the next 14 days on Arena, GPT-Image-2.5 Sunburst by @OpenAI will be available to test directly on Arena!

@OpenAI’s latest image model is ranked #1 in the Text-to-Image Arena, the Image Edit Arena, and in the Multi-Image Edit Arena. Find the link below to Direct Mode, where you can now select GPT-Image-2.5 Sunburst from the dropdown menu.

Access to GPT-Image-2.5 Sunburst ends November 2nd at 8 a.m. PT. Afterward, it will remain available anonymously in Battle and Agent Mode. User votes from Battle and Agent Mode are used to power our real-world leaderboard rankings. In these modes, models are always anonymized.

[RTK @T2IRedTeamKing]

Now that I think about it in terms of Direct Chat, this might be the same fate for the what-if Nano Banana Pro 3 (let’s assume between Nov and Dec 2026) on Arena AI. Back then, it wasn’t really like this.

[RTK @T2IRedTeamKing]

You hit the nail on the head. They love the aesthetic of a technical benchmark, but as soon as someone asks a basic common-sense question, the playbook for Arena AI falls apart. They are too busy with planning their next rushed marketing.

[David Hampl @davhampl]

@arena How do you score "unauthorized action" when the agent's whole job is acting? 90K sessions of what, exactly?