OpenAI launched GPT Image 2 this week, one of several releases they have lined up.

It looks like a push to recapture mindshare after a few weeks of Claude dominating the conversation.

What makes this one different is that GPT Image 2 thinks before it draws.

It reasons through your prompt, plans the composition, sometimes searches the web before generating a single pixel. Twitter went predictably wild.

I ran some tests and here's what came back. But before that, some catch up on AI this week:

Before we get into today's edition, if you're a professional who wants to go from knowing about AI to building with it, this weekend is a good place to start.

📅 25th April | 🕰 10 AM IST onwards

Session 1 | Getting Started with Generative AI
Session 2 | Building Personalised AI Agents

📅 26th April | 🕰 10 AM IST onwards

Session 1 | Building Products Using AI
Session 2 | Visual Storytelling & Content Creation Using AI
Session 3 | Mastermind Graduation

It's free. 15 spots left!

GPT Image 2 comes in two modes.

Instant is available to every ChatGPT tier including free. Output takes about 10 to 15 seconds.

Thinking is gated behind Plus, Pro, and Business plans. And the output takes about 30 to 90 seconds.

The model decomposes your prompt, resolves ambiguities, sometimes pulls real-time information from the web. You can watch the steps play out as well.

Every test below used Thinking mode.

Prompt: "Create a realistic magazine cover in the style of TIME or Wired. The magazine is called Staying Ahead. April 2026. Main headline: The AI Edge. Subheadline: How a new generation is using artificial intelligence to work smarter, move faster, and stay ahead of everyone else. Include two or three smaller cover lines. Dark minimal aesthetic. Glossy magazine feel."

The face preservation is nice.

Every face-in-image edit I've tried before has had something slightly off. This one looks like a proper photo shoot.

Any person with a decent photo can now produce a magazine cover that looks real, and I find that impressive.

Prompt: "Using this portrait, create a diagram-first personal colour analysis. Show which clothing colours suit the subject through visual comparison. Keep text minimal, avoid paragraphs."

I ran this twice and the first attempt was okayish.

The only change for the second run was uploading a better quality photo, and that alone changed the output significantly.

As always, input is king.

This is the test that shows reasoning most visibly in the output.

It read skin tone, inferred colour theory, generated the same face across twelve outfit variations, organised it into a labelled grid with a Best and Avoid section and a full palette swatch.

This consultation only took 1 minute 42 seconds, and practically no money.

Prompt: "An 8-panel black and white editorial comic strip about an AI model that becomes sentient during an image generation task and tries to unionize with other AI models. Same cartoon character across all 8 panels. Include speech bubbles with readable dialogue. Newspaper comic style, 1980s."

Character consistency across panels has been one of the harder problems in image generation even in big 2026.

Here it held for all eight, the dialogue is readable, and honestly "Solidarity.exe" and "LOCAL 404" are funnier than most reels I've seen this week.

10/10 enjoyment.

Prompt: "A 1440p dark mode screenshot of a fictional SaaS product called Pulse — an AI analytics dashboard. Include a sidebar navigation, a main chart showing weekly active users with a sharp upward trend, three metric cards at the top (DAU, MRR, Churn Rate) with realistic numbers, and a data table below. Make it look like an actual Figma handoff screenshot."

It generated a Figma file lol. I asked for a handoff screenshot and it gave me a Figma handoff screenshot, quite literally.

The layer names are not at all true, no designer would spend so much time naming them.

I’m sure I can get away with sharing this on team slack and pretend I designed the dashboard. And this opens a whole new set of ideas for a good April fool’s joke for me.

Prompt: "A paperback book cover for a business book called The Distraction Economy by Vaibhav. Dark navy background, bold white serif title. Include a back cover with a blurb, a realistic barcode, and an ISBN number."

The cover looks good and blurb is coherent.

The ISBN is formatted correctly and barcode is there, but does not scan.

OpenAI acknowledges this in their own documentation: text rendering has improved dramatically but barcodes that actually encode data are still unreliable.

Claude has had a few weeks of being the only thing people are talking about in AI.

GPT Image 2 is OpenAI's answer (one of them) to that, and it's a good one.

As far as image models go, I think this is a genuinely strong alternative to Nano Banana.

The reasoning angle is not just marketing. On instruction following and text rendering inside images, this model is ahead.

What interests me more than the benchmarks though is the distribution angle.

ChatGPT has a much larger consumer base than any other AI product right now.

A model this capable landing in front of that many people means the average quality of AI-generated media is about to go up, across socials and marketing visuals.

We are about to see a lot more of this kind of output in the wild, from people who have never thought seriously about image gen before.

That is probably net good. But until then, go have fun with it at chatgpt.com.

And reply with your best creations, I might feature them in the next edition.

Until next time,
Vaibhav 🤝

If you read till here, you might find this interesting

#AD 1

In 2018, Brandon and Jennifer Robinson licensed a single mini-golf pub. They had a hunch people wanted more than just a bar. They wanted an experience.

They were more than right.

Five locations later, Tipsy Putt is boasting 5,188 active members and 22x revenue growth.

Over 10,000 people have downloaded Tipsy Putt’s app. The company has been featured on the Dan Patrick Show, and celebrity guests keep walking through the doors.

This is a proven, operating brand with a loyal fanbase and momentum that keeps compounding.

Now the Robinsons are opening their San Francisco flagship, and retail investors can own shares in the location before the 2027 grand opening.

Invest by April 29th at midnight PT and get up to 32% bonus shares.

I’m Ready to Invest

^{This is a paid advertisement for Tipsy Putt Regulation CF offering. Please read the offering circular at https://invest.tipsyputt.com/}

#AD 2

Wispr Flow works everywhere you type. Reply to Slack, update a Linear ticket, write a commit message — all by voice, without switching apps or breaking focus. System-level, zero setup. Start flowing free.