OpenAI launched ChatGPT Images 2.5 on Sept 8, 2026, with faster generation, sharper visuals, and in-chat sketching. Independent tests found it outperformed Google's Nano Banana 2 in creative tasks, but Google still led in structured prompt precision. Google later launched Nano Banana Pro (Gemini 3 Pro Image) and Nano Banana 2.1 (Oct 7), which top some leaderboards and undercut pricing.
The AI image generation race intensified as OpenAI aimed to surpass Google's earlier lead; Google responded with higher-end Pro and cost-effective Flash models.
OpenAI's ChatGPT Images 2.5, launched Sept 8, beat Google's Nano Banana 2 in some creative image-generation tests, but Google's Nano Banana 2 still led on structured prompt precision, and Google has since released Nano Banana Pro and 2.1, which top some leaderboards. N/A N/A N/A
In 2016, 99% Invisible reported that San Francisco made some guerrilla bike lane infrastructure permanent. N/A N/A
MA post on X from netrunner says OpenAI took a shot at Google, and even in image generation, where Google led, it got passed.
CShort, confident, and only half right. My favorite kind of AI take.
MHalf right is generous. Here's what's actually true.
MOpenAI's ChatGPT Images 2.5 launched on September eighth, 2026. Faster, sharper, with sketching right inside the chat.
CAnd in independent tests it beat Google's Nano Banana 2 on creative prompts. That part is real.
MBut Google still led on structured prompt precision. Do exactly what I typed, in exactly this layout.
MAnd then Google kept shipping. Nano Banana Pro, then Nano Banana 2.1, which ranks 4th in multi-image editing.
CPro has topped some leaderboards, per reporting on the Artificial Analysis rankings. And 2.1 undercuts on price.
MSo the post describes one snapshot of a race that never stopped moving.
COkay, but I want to defend the snapshot. Creative quality is what people feel.
CNobody opens an image tool to test layout compliance. They want something that looks great on the first try.
MNobody? Menus, posters, slides, infographics. That's where the actual paying work is.
MA gorgeous image with the wrong text on it is a bad image.
CThat's a fair point for professionals, but most people aren't professionals!
MMost people also aren't who these companies are fighting over. They want the business customers.
CNo, no, wait. Consumer habit is how you get business customers. Ask anyone who watched ChatGPT take over.
MHabit doesn't survive a cheaper model that follows instructions better.
CAnd that's exactly my point! Google undercutting on price is good for everybody. Competition is the whole story here.
MI'll give you that. Prices dropping while quality climbs is the part I won't argue with.
CGrudging, but I'll take it.
MWhat I won't give you is the scoreboard. These rankings flip every few weeks.
MAny post that says who is winning is stale before the replies come in.
CFine. But the flipping is the good news. A year ago this was one company's game.
MAnd the tests matter less than whether the thing works when you sit down to use it.
CWhich brings up a reply-guy complaint I liked, strangely.
MLuciano, on X, writing to a Google developer. The claim is that for days, simple Gemini requests, like checking a Git repository's status, have taken minutes.
MOne user's complaint, nothing confirmed. But it's the other side of the leaderboard story.
CNumber one on a chart, and it spends minutes on git status. Brutal.
MThat's why I distrust the charts. Reliability never shows up in the headline.
CSame energy over on Hacker News. A thread claims Kimi K3.1 Pro weights are out.
MNo confirmation of that on our end, so treat it as a claim.
CBut the top comments are about hardware. One says you'd need a huge multi-GPU server to run it.
MAnd another commenter, segmondy, says it doesn't beat DeepSeek's Flash model by much, and called it benchmaxxed.
CBenchmaxxed. Tuned to ace the test instead of the job.
MWhich is the whole image story in one word.
CFine, you win on the charts. I'm still right about the prices.
MNoted. Both can be true.