On Tuesday, OpenAI quietly launched ChatGPT Images 2.0. The API name is gpt-image-2. It's the first image model from OpenAI that reasons before it draws. And it's killing DALL-E — 2 and 3 are both retired on May 12.

I spent the last two days running this thing through every aesthetic I could think of, and every image in this newsletter was made with it. The fashion editorial Petra Collins avatars in the Business Idea section. The Andreas Gursky mega-scale data center shots. The Rockwell-gone-wrong small-town vignettes. All gpt-image-2.

Here's what actually changed.

It thinks first. The model has a Thinking mode (Plus, Pro, Business) that reasons through the prompt before generating. It can also search the web mid-generation to ground visual details in reality. Ask it for "Jensen Huang at Computex 2026 in his signature leather jacket" and it'll pull the right jacket. It verifies its own output before handing it to you.

Eight images at once, with consistency. One prompt, up to eight outputs, and the same character shows up across every one. VentureBeat covered this angle: you can now generate an entire manga sequence, a children's book, or a brand's full Instagram grid with a unified visual DNA. Character sheets with the same face from multiple angles. Product shots across ten contexts.

Text actually works. Multilingual too. Japanese, Korean, Chinese, Hindi, Bengali — it renders them all. The "AI can't do text" era is officially over.

2K resolution. Free tier gets Instant mode. No paywall on the base model. You can play with it today for zero dollars.

It landed at #1 on Image Arena with a 242-point lead across every category. Gemini held the top spot a week ago. Not anymore.

Here's why this matters beyond pretty pictures. The consistency thing is the unlock. Every AI image tool until this week made one good image, then shrugged when you asked for the same character in a different scene. That broke every commercial use case — storyboards, e-commerce lifestyle shots, ad variants, avatars. Now? You prompt once, get eight matching variants, and ship them. That's the difference between a toy and a production tool.

If you're building anything visual — a newsletter, a Shopify store, a podcast, a brand — your content pipeline just collapsed by an order of magnitude. And if you're building AI influencers? See the Business Idea section below. This model just made that play 10x easier.

You can now buy a humanoid robot for $4,500. Unitree's R1 is shipping this month. The company moved 5,500 units in 2025 and is targeting 20,000 this year — more than Tesla, Figure, and Agility combined. A walking robot that costs less than a used Civic. We are here.

ChatGPT just landed in CarPlay. You can now ask it questions straight from your dashboard, voice-first, hands-free. Every commute just got weirder.

NASA unveiled the Nancy Grace Roman Space Telescope at Goddard this week. It's Hubble's successor with a field of view 100 times larger. Launching in 2027. It's going to map dark energy in ways we've never been able to see. Space nerds, your time has come.

Google rolled out personalized image generation in Gemini, pulling from your own Google Photos library via Nano Banana 2. You can generate pictures of yourself, your dog, your kids in scenes that never happened. That's either the coolest thing that shipped this week or the most quietly terrifying. Probably both.

Huawei launched AI eyewear alongside a new foldable, a smartwatch, and the Pura 90 phone. The Meta Ray-Bans finally have real competition. The "Huawei is dead" narrative is aging badly.

Anthropic started requiring government photo IDs and selfies from some users to block access from China, Russia, and North Korea. A first for a frontier lab. Meanwhile Trump told CNBC a DoD deal with Anthropic is "possible". The frontier labs are officially becoming defense contractors. Take the screenshot now.

Apple confirmed this week that Tim Cook will become executive chairman effective September 1. The keys go to 50-year-old John Ternus, longtime SVP of Hardware Engineering. He's been at Apple 25 years. That's literally half his life. He's the guy behind Apple Silicon, the M-series, and Vision Pro.

Cook's 15 years were historic. Apple's market cap went from around $350B to roughly $3.5 trillion on his watch.

Here's what I keep chewing on. Apple is handing the company to a hardware engineer at the exact moment AI is eating software. That's either the smartest move of the decade (because Apple's real moat is silicon) or the worst-timed handoff imaginable (if agentic is the future and the device stops mattering). The 2027 iPhone event tells us which one.

Jeff Bezos is nearing a $10 billion funding round for his stealth AI startup Project Prometheus, now valued at $38 billion per the FT. JPMorgan and BlackRock are leading. The company launched quietly in November 2025 with $6.2 billion.

The pitch isn't chatbots. Prometheus is training AI systems that understand physics. Think less "write me a poem" and more "redesign this jet engine with 12% less drag."

Here's the play nobody's ranking on yet. Bezos is reportedly also raising up to $100 billion for a separate holding company to acquire industrial businesses and feed their operational data back into Prometheus. Buy the factories. Feed the models. Own the loop.

Bezos isn't trying to out-OpenAI OpenAI. He's betting the trillion-dollar prize is the physical world, not the chat box. If he's right, Anthropic and OpenAI are fighting over a tiny market.

OpenAI dropped its biggest Codex update yet this week. Codex can now run your computer. It sees, clicks, types. It plugs into 90+ apps (Atlassian, GitLab, Microsoft Suite), remembers your preferences, runs scheduled automations. Over 3 million developers use it weekly.

In the same 24 hours, OpenAI quietly turned on cost-per-click ads inside ChatGPT. Bids are $3–5 per click. That's on top of the CPM ads they launched earlier.

Codex just stopped being a coding tool. It's a general-purpose computer agent, aimed squarely at Anthropic's Claude Code playbook. And CPC ads inside ChatGPT means answer quality will start depending on what advertisers paid. Set a reminder for when "sponsored" results start appearing in your responses. Because they will.

Look at what shipped in the last five days. A humanoid robot you can put in your living room for $4,500. ChatGPT in your car. Huawei AI eyewear on your face. A Bezos-funded lab whose entire thesis is "AI that redesigns jet engines." A NASA telescope that'll map dark energy.

The chat window is the AI equivalent of a command-line interface. It's where we are because we don't know where we're going yet. The companies figuring out the physical interface — robots, cars, glasses, factories — are going to eat the ones still obsessing over model benchmarks. Watch what Bezos is doing. Watch what Unitree is doing. Watch what Huawei is doing. The labs are late.

Apple spent two years getting dragged for being behind on AI. Ternus's promotion tells you exactly where Tim thinks the moat lives. The deal to run a 1.2T-parameter Gemini on Apple's Private Cloud Compute is already done. Apple doesn't want to build the brain. Apple wants to be the trusted hardware layer every AI runs on top of. Ternus's whole job is keeping that layer uncatchable.

People calling Apple doomed are playing the wrong game.

You were warned. The moment ChatGPT started serving ads at CPC rates, the incentive structure flipped. OpenAI's KPI used to be "answer questions well." Now it's "maximize time on site and ad-eligible queries." Same arc Google took from 1998 to 2018, compressed into 24 months.

ChatGPT's answers will get subtly worse over the next year and a half. Not because the models got worse. Because the product got optimized. If you've built anything on ChatGPT's surface, start building loyalty on your own channel now. Email list. App. Anything you own.

Here's the gap. Brands spent over $21 billion on influencer marketing in 2024, and most of it goes to UGC creators charging $300–$1,500 per 30-second video. For a small DTC brand that needs 20 pieces of content a month, that's $10K+ before anyone has clicked "add to cart."

Last week I spun up a photorealistic AI avatar on Higgsfield in under ten minutes and generated a social proof video for my sports podcast merch. I genuinely could not tell it wasn't a real person. Neither could anyone I showed it to. The tech quietly crossed the uncanny valley and nobody sent out a press release.

The play. Become the AI influencer shop for one niche. Supplement brands. Coffee roasters. Skincare. Pet products. Pick one. Build 5–10 reusable AI avatars with distinct vibes — the gym bro, the skeptical mom, the college kid, the corporate professional. Use gpt-image-2 (the model I was just raving about) to generate consistent character sheets — eight angles of the same face, product in hand, different backdrops. Then drop those into Higgsfield Marketing Studio and animate them into 30-second UGC-style videos. Charge $500 per video, or $3,000/month for a bank of 20. Sign 5 brands on retainer and you're at $15K/month working 20 hours a week.

How to do it. Week one, build the avatar library. Use gpt-image-2 to lock in 10 consistent characters (save every prompt), then port them into Higgsfield for motion. Week two, productize. Pick a niche and build a repeatable script template: hook, product beat, social proof line, CTA. Week three, cold-DM 30 small Shopify brands and offer three free videos in exchange for a testimonial. Week four, use the case studies to close paying retainers.

Why it works. Real UGC creators have two constraints. They cost money and they have schedules. AI avatars have neither. You can ship 50 videos in a day, A/B test hooks at scale, and localize for different markets without flying anyone anywhere. The moat isn't the tech — anyone can open gpt-image-2 and Higgsfield. The moat is the prompt library, the hook templates, and knowing which avatar converts for which niche. Knowing which of ten avatars sells protein powder to 28-year-old guys in Texas is the part nobody can Google.

No affiliate angle? Same playbook, just point the AI avatars at an Amazon Associates or Impact store instead of a client. Same content engine, different monetization. One friend of mine is quietly making four figures a week doing exactly this on a finance niche account.

One caveat. The FTC has updated its endorsement guides to cover AI-generated talent. Disclose it in the caption or on-screen. Brands that skip disclosure are going to eat fines. Brands that disclose cleanly are going to eat everyone else's lunch.