The Week AI Learned to Draw Text

The bakery is called Marigold. It's on a corner in Greenpoint, Brooklyn, and last Tuesday its owner, who has never opened Photoshop in her life, needed a sign for the front window. Not a logo. Not a brand identity. A hand-lettered-looking chalkboard sign that said "SOURDOUGH BOULES — BACK SATURDAY" in a font that didn't look like it was generated by a computer in 1997.

She typed that sentence into ChatGPT, described the chalkboard she wanted, and hit generate.

The sign came back with every letter legible. Every word spelled correctly. The chalk texture looked like someone had actually dragged a stick of calcium sulfate across a dark green surface.

That has never happened before.

What actually shipped on April 21

ChatGPT Images 2.0 launched on April 21, 2026, and the specific thing it does that nothing before it could do reliably is render text inside images. Not garbled text. Not "close enough if you squint." Actual readable words, in the right order, spelled correctly, placed where you asked.

The technical specs: 2K resolution output, aspect ratios ranging from 3:1 panoramic all the way down to 1:3 portrait. Up to eight coherent images from a single prompt, with character and object continuity across the set — meaning you can ask for a series of event flyers with the same mascot in different poses and get something usable instead of eight unrelated hallucinations.

There's also a thinking mode that enables web search, layout reasoning, multi-image batching, and output verification. That last part matters: the model checks its own work before handing it back. Thinking mode is restricted to Plus ($20/month), Pro ($200/month), Business, and Enterprise tiers. Free users get the base generator, which is still a significant leap from what existed a month ago, but without the self-verification loop.

And it handles non-Latin scripts — Japanese, Korean, Hindi, Bengali — with the same reliability. That detail alone opens the tool to roughly four billion people who were effectively locked out of every previous AI image generator's text rendering.

What it actually does well

Let me be specific, because "it renders text" is the headline and the reality is more nuanced than the headline.

Short text on designed backgrounds. A poster title. A book cover with an author name. A product label. A business card. A social media graphic with a quote. Anything where the text is fewer than about forty words and the design context is clear, ChatGPT Images 2.0 handles with startling competence. You describe the aesthetic — minimalist, vintage, hand-drawn, corporate, playful — and the text sits inside that aesthetic like it belongs there.

Multi-image consistency. This is the feature that actually changes workflows. Ask for a set of five Instagram story slides with the same visual language and the same character, and you get five images that look like they came from the same designer in the same afternoon. The continuity isn't perfect — hair color might shift a shade between frames, background elements might migrate — but the overall coherence is miles beyond what any image generator produced even six months ago.

Mockups and prototypes. A student building a presentation deck. A small business owner mocking up a menu. A freelancer pitching a client on a visual concept before committing to the real design. These are the use cases where "good enough to communicate the idea" is the entire point, and ChatGPT Images 2.0 is genuinely good enough to communicate the idea.

Signage and physical-world text. Restaurant specials boards. Event banners. Yard sale posters. The kind of thing where you used to open Canva, fight with text boxes for twenty minutes, and end up with something that looked like a Canva template. Now you describe what you want and get something that looks like a person made it.

What it still gets wrong

Here's the part where I tell you what the breathless launch coverage left out.

Long-form text is still unreliable. Ask it to generate a full page of body copy — a flyer with three paragraphs of event details, say — and the coherence starts to buckle around the second paragraph. Letters swap. Words merge. The spacing between lines goes uneven. If your design needs more than about fifty words of readable text, you're still better off generating the visual separately and adding the text in PPhotopea MCP Server or any basic editor.

Fonts are suggestions, not specifications. You can ask for Garamond. You will not get Garamond. You will get something that has the general vibe of a serif typeface with oldstyle proportions, and that might be fine for your purposes, but if you need brand-specific typography, this tool is not your endpoint. It's your starting point — generate the layout, then swap the fonts in a real editor.

Text placement precision is approximate. "Put the title in the upper left corner" gets the text in the upper-left-ish region. "Align the subtitle exactly 40 pixels below the title" is a request the model understands conceptually but cannot execute with pixel-level precision. For social media graphics where "close enough" is close enough, this is fine. For print design where registration matters, it's not.

It still hallucinates decorative text. Background elements that look like they contain writing sometimes do contain writing — and it's gibberish. A bookshelf in the background might have "book titles" that are nonsense words. A storefront sign in the distance might say something that isn't a real word. The main text you asked for is clean; the ambient text the model invents on its own is still the old chaos.

Complex layouts with multiple text blocks at different sizes. A concert poster with a headline, three support acts, a date, a venue, and a ticket price. That's six distinct text elements at different scales, and the model starts dropping or garbling the smaller ones. The headline will be fine. The headliner name will be fine. The venue name in 10-point type at the bottom will be a coin flip.

The graphic design conversation

Creative Bloq ran a headline about people declaring the death of graphic design. Again. The "again" is doing a lot of work in that sentence, and it's the right word.

Here's what's true: a person who has never touched a design tool can now produce visual assets that would have required either hiring a designer or spending three evenings learning Canva. That's real. That changes who gets to make things, and that matters.

Here's what's also true: the gap between "produced by someone with no design training using an AI generator" and "produced by a designer who understands typography, hierarchy, color theory, whitespace, and the specific emotional weight of a particular typeface at a particular size" is not a gap that AI closed this week. It's a gap that AI made more visible, because now non-designers can get 70% of the way there in thirty seconds, and the remaining 30% is where design skill actually lives.

The bakery owner doesn't need the remaining 30%. Her chalkboard sign is going in a window on a street where people are walking past at three miles an hour. It needs to be legible and attractive. It does not need to be Pentagram.

A startup raising a Series A needs the remaining 30%. Their pitch deck needs to look like a company that pays attention to detail, because investors read presentation design as a proxy for operational discipline. ChatGPT Images 2.0 can mock up that deck, but it can't finish it.

A freelance designer should be using this tool. Not because it replaces their work — it doesn't — but because it collapses the concepting phase. Instead of sketching four directions by hand and presenting them to a client, a designer can generate twenty directions in ten minutes, curate the best four, and present those. The client conversation starts farther along. The final product still requires the designer's hand.

That's not the death of graphic design. That's a power tool. Carpenters weren't replaced by nail guns. They were freed from spending their wrists on repetitive hammering so they could focus on joinery.

What a non-designer can actually make on a Monday morning

Let me walk through five realistic things a real person with no design background could produce before their second cup of coffee using ChatGPT Images 2.0 and a Plus subscription.

A social media announcement. "We're hiring a part-time barista. Apply in store." On a warm-toned background with your shop's name visible. One prompt, one generation, one post to Instagram. Done in under two minutes.

A birthday party invitation. "Dinosaur-themed party for a 6-year-old named Marcus. Saturday June 7, 2-4pm, Riverside Park Pavilion B." With cartoon dinosaurs and the text laid out like a real invitation. You'll probably want to generate two or three versions and pick the best layout, but the whole process takes five minutes.

A presentation title slide. You're giving a talk at a PTA meeting about the school's reading program. You need one strong visual with the program name and the date. That slide used to come from a PowerPoint template that looked like every other PowerPoint template. Now it looks like someone designed it.

A product label. You make candles in your garage and sell them at the farmers market. Each scent needs a label. "Cedarwood & Rain, 8 oz, hand-poured soy candle" on a label that looks like it came from a small-batch artisan brand instead of a laser printer. You can generate a set of eight labels for your full product line from a single prompt, and they'll share a visual language.

A flyer for a lost dog. This is the use case nobody talks about, but it's the one that matters most on the worst Monday morning of your life. You need a picture of your dog, a phone number, and the word LOST in big letters. ChatGPT Images 2.0 won't use a photo of your actual dog — that's a limitation; see Your AI Can't Draw Your Dog (Yet) for why — but it can generate a breed-accurate illustration with all the right text, legible from across a street, in under a minute.

The pricing question

Thinking mode — the one with self-verification, web search, and multi-image batching — requires a Plus subscription at minimum. That's $20 a month. For a small business owner who would otherwise spend $50-$200 per design project on Fiverr, the math works immediately. For a student who just needs occasional graphics for presentations, it's a tougher call — but the base generator on the free tier still produces dramatically better text rendering than anything that existed last month.

The Pro tier at $200/month only makes sense if you're generating high volumes daily — a content creator, a marketing team, an agency using it as a concepting engine. For everyone else, Plus is the sweet spot.

The non-Latin breakthrough nobody covered

The English-language tech press moved past this detail in a single bullet point, but it deserves its own section. ChatGPT Images 2.0 renders text in Japanese, Korean, Hindi, and Bengali with the same reliability it handles English. Previous generators treated non-Latin scripts as decorative — they'd produce something that looked vaguely like Japanese calligraphy but was, upon inspection by an actual Japanese reader, nonsense. Beautiful nonsense, but nonsense.

That's over. A shop owner in Mumbai can generate a Hindi-language storefront banner. A Korean student can make a presentation title card in Hangul. A Bengali poet can produce a book cover with their own lines rendered accurately. These aren't edge cases. They're the majority of the world's literate population, and they've been locked out of usable AI image generation since the category existed.

The technical reason this works is that thinking mode's output verification step includes script-aware text checking — the model doesn't just verify that text was placed, it verifies that the characters are correct and properly formed. This is a hard problem in generative image models because non-Latin scripts often have complex character composition rules (ligatures in Hindi, stroke order significance in Japanese kanji), and getting them wrong produces output that's immediately recognizable as machine-garbled to a native reader.

ChatGPT Images 2.0 doesn't solve this perfectly — complex compound characters in Bengali still occasionally render with spacing issues, and very small kanji can lose detail — but it crosses the threshold from "unusable" to "usable with minor touch-ups." That threshold is the one that matters.

How it fits with the tools you already know

If you're already using AI image tools, ChatGPT Images 2.0 doesn't replace them so much as it fills the specific gap they all had. AAI Creator is still excellent for workflow automation and batch generation. OOpenAI GPT Image MCP lets developers integrate the same underlying model into their own pipelines. The difference is that ChatGPT Images 2.0 is the consumer-facing version — no configuration, no API keys, no code. You type what you want and you get an image with readable text.

For compositing and detailed editing after generation, PPhotopea MCP Server remains the right tool. Generate the base image in ChatGPT, bring it into Photopea for precise text adjustments and layer work. That two-step workflow is going to become very common very quickly.

The prompt that actually works

After a week of testing, here's the structure that produces the most reliable results for text-heavy images:

Describe the physical format first (poster, business card, label, social media graphic). Then describe the visual style (minimalist, vintage, illustrated, photorealistic). Then state the exact text you want rendered, in quotation marks, with line breaks indicated. Then describe the color palette or mood.

The quotation marks matter. They tell the model "this is the literal text, not a description of what the text should be about." Without them, you get paraphrased versions of what you wrote. With them, you get the words you asked for.

If you want to try the specific prompt structure that produced the best results in testing for business graphics, The Five-Minute Logo walks through it step by step with a copy-pasteable template.

What this week actually means

The chalkboard sign in the window of Marigold bakery in Greenpoint is not the end of graphic design. It's the beginning of a world where a person who makes bread for a living doesn't have to also learn design software to tell people the sourdough boules are back on Saturday.

That sounds small. It isn't.

Every gatekept skill that gets automated doesn't just save time for the person who needed it — it returns attention to the thing they actually do. The baker bakes. The student studies. The freelancer focuses on the work, not the marketing collateral for the work.

ChatGPT Images 2.0 is the first AI image tool that handles the most common reason a non-designer opens a design tool: they need an image with words on it. The words needed to be right. For the first time, reliably, they are.

The sourdough boules are back on Saturday. The sign says so, and you can read it from across the street.