Introducing ChatGPT Images 2.0
TL;DR
OpenAI’s new image model looks like a real leap, not a cosmetic update — Wes highlights the jump on the Artificial Analysis leaderboard from Gemini 3.1 Flash Image Preview’s 1270 to GPT Image 2’s 1512, calling it a “massive leap” over prior image systems.
The core pitch is interactive image intelligence, not one-shot generation — in OpenAI’s launch demo, the team repeatedly framed Images 2.0 as something you “talk to,” with follow-up edits, multi-image coherence, web search, QR code generation, and a new “thinking mode” before rendering.
Text rendering is the breakout capability people will actually notice immediately — the demo showed dense, legible posters in Japanese, Hindi, Chinese, and other languages, plus absurd precision examples like “GPT image” written on a single grain of rice in a 4K experimental image.
Wes and Dylan found the editing workflow promising but still quirky in real use — transparent PNG backgrounds worked, iterative edits preserved object layout surprisingly well, but things broke under repeated generations, like transparency turning into checkerboard artifacts and prompts drifting in odd ways.
It feels smarter than older image models, but not reliably literal — their stress tests showed both the upside and the limits: a blue pen labeled “red” worked, but a “glass of red wine filled to the brim” kept turning into half-full wine or a weird goblet/beer-glass hybrid.
The most interesting gap is between ‘technical correctness’ and human commonsense — Wes’s Where’s Waldo prompt got interpreted as “make Waldo hard to see,” so the model turned him semi-transparent instead of hiding him naturally in a crowd, which perfectly captures where this new reasoning still misses human intent.
The Breakdown
A chaotic livestream start, then straight into OpenAI’s launch
Wes and Dylan open in full scrappy livestream mode: dual-stream confusion, privacy settings on private, and jokes about bad hair visibility. Once they get YouTube behaving, they jump into OpenAI’s live reveal of ChatGPT Images 2.0 without pretending this is polished TV.
OpenAI’s first big claim: images you can actually talk to
The launch demo starts with fashion try-on style prompting: zoom in, show alternate views, preserve clothing details, and keep text labels coherent. The key line from OpenAI is that this is “no more like an AI image generator” that returns one picture, but an image system you interact with conversationally.
“Thinking mode” and why OpenAI says this is different
Kenji introduces a deeper capability: the model can think before producing the final image, which OpenAI says helps with complex prompts, consistency across multiple outputs, web search, and self-checking. Their examples are very showy — a manga sequence of Gabe and Sam that stays visually consistent across pages, plus an image that quotes reactions from Threads, LinkedIn, and Reddit and even embeds a working QR code to ChatGPT.
Photorealism, 360 scenes, and weirdly strong spatial coherence
Alex shows how prompts like “photorealistic,” “shot on iPhone,” or “disposable camera” trigger much more natural-looking output, including a faux-2015 OpenAI lecture scene with coherent slide text. The standout moment is a 360-degree moon landing panorama that they drop into a custom viewer; even Wes sounds genuinely impressed that the shadows and sun direction hold together.
The text rendering flex: Japanese posters, Hindi recipes, and rice-grain typography
This is where the demo really lands. OpenAI’s team shows multilingual poster generation in Japanese, Chinese, French, and Hindi, stressing that the model now handles dense scripts with thousands of characters much better than before. The biggest crowd-pleaser is the experimental 4K image of a rice pile with “GPT image” written on a single grain — Wes calls that moment mind-blowing.
Wes and Dylan start testing it themselves — and the first results are strong
Back on their stream, they immediately riff on OpenAI’s goofy long-neck portrait style and test the model on personalized generations. Wes pulls up examples including him and Sam Altman playing Nintendo, a fake old-time newspaper about Tim Cook leaving Apple, and a modern Einstein classroom scene, mostly reacting to how shockingly readable the text is.
Real-world editing is where the model gets interesting — and messy
The pair stop caring about pretty pictures and start probing usability: can it act more like Photoshop? Dylan generates bizarre food combos, swaps in a blue pen labeled “red,” and gets a genuinely transparent PNG background working, which they both see as hugely useful for thumbnails and design workflows. But repeated edits start to reveal cracks: checkerboard transparency artifacts get “burned into” later generations, and running too many tabs at once seems to make outputs glitch.
The funniest tests expose the model’s remaining blind spots
Their best prompts are basically traps for human common sense. A “glass of red wine filled to the brim” repeatedly fails in spirit even when it sort of succeeds technically, and a Where’s Waldo prompt becomes “Predator Waldo” because the model makes him translucent instead of cleverly hidden. By the end, Wes’s verdict is still positive — the jump feels real, the intelligence is deeper, and “thinking before drawing” matters — but the weird misses are exactly what make the session fun to watch.