Where image generation actually stands across the major labs, April 2026

OpenAI shipped ChatGPT Images 2.0 this morning. Before anyone had time to write a take, I ran five prompts through five models and scored every axis. GPT Image 2 won 25/25. That is today's story.

The bigger story is that image generation is now a stack of specialists. Each major lab owns a specific job. Picking the wrong one either burns cost or ships a hero that does not land.

Here is the honest map as of April 21, 2026. The first two entries are the models I ran today. The rest are positions in the field I have opinions on from ongoing use, where I have noted what I have directly tested versus where I am leaning on their publicly-stated positioning. Check current versions and pricing before you commit to anything.

1. OpenAI GPT Image 2 (shipped today)

What it does well: UI mocks with real-looking text, product screenshots of things that do not exist yet, multi-subject editorial scenes with internally consistent detail. It was the only model in my launch-day test that drew a complete, correctly-wired microservices diagram on a whiteboard when asked.

What it costs: $0 marginal via ChatGPT Plus today (Thinking mode on). API access via gpt-image-2 requires OpenAI org verification through Stripe Identity, typically a 1 to 3 day process. Once through, expect pricing in the same ballpark as other frontier image APIs.

The specific capability that shifted today: UI realism. Before this morning, Google's Nano Banana Pro was the model I reached for when I wanted a ChatGPT macOS window that looked real. After this morning, GPT Image 2 is. That swap happened in a single release.

Where it falls short: not yet broadly indemnified for enterprise use. Weaker on vector output than Recraft. Weaker on dense logo-quality typography than Ideogram.

2. Google Nano Banana Pro and Nano Banana 2 (Gemini 3 Pro Image + 3.1 Flash Image)

What they do well: both tied at 23/25 on my same-morning test. Pro is more literal, the Flash variant is more atmospheric and cheaper.

What they cost: roughly $0.134 per image for Pro and $0.067 for the Flash variant, both via Vertex AI at the pricing I saw in my run today. The Flash variant is where you run bulk automation that does not need to ship as a hero.

The specific capability that still matters: API access with no verification gate. gemini-3-pro-image-preview and gemini-3.1-flash-image-preview on Vertex take a straight generateContent call with responseModalities: ["IMAGE"]. Works today, no queue. If you need images at API velocity before OpenAI's verification clears, this is the path.

Where they fall short: the UI-realism gap to GPT Image 2 opened up today. Multi-subject scenes with text consistency still trail.

3. Black Forest Labs FLUX (open-weights leader)

What it does well: the strongest open-weights family of models at the frontier. Stronger per-pixel fidelity than any other open-weights option. Reference-image conditioning for character and brand consistency is the specific thing that has pulled Midjourney power users across.

What it costs: free if you run it yourself. The larger variants want a serious GPU; via Replicate or similar hosted endpoints you are in the low cents-per-image range. I have not re-tested current pricing this week, so verify before committing.

The specific capability that matters: you own the weights. No API gate, no T&C surprise, no "we ran out of capacity." If your product needs image generation to run reproducibly for years, or your legal posture requires self-hosting, FLUX is the default.

Where it falls short: still behind GPT Image 2 and Nano Banana Pro on real-product-mock tasks with embedded text. Good for concept art, illustration, photorealistic composites. Not the first reach for landing-page heroes.

4. Midjourney (best editorial aesthetic)

What it does well: the aesthetic ceiling. Editorial-quality photography, cinematic scenes, portraits that look like high-end ad campaigns rather than AI renders. Style consistency and prompt adherence continue to improve without losing the house look.

What it costs: subscription tiers on Discord + web starting in the tens of dollars per month. No public API; programmatic access requires unofficial channels.

The specific capability that matters: "make this look like it came from a real magazine shoot." For marketing imagery, merchandise mockups, branded photography, Midjourney still wins on aesthetic sense of place.

Where it falls short: the missing public API kills it as a production dependency. Weaker than the frontier on text rendering and real UI mocks. Web app is still the cleanest surface.

5. Ideogram (dense typography specialist)

What it does well: text. Logo-quality type. Dense packed words on posters, merch designs, editorial layouts. The only model I trust to render a twenty-word quote on a hand-lettered poster in one pass.

What it costs: subscription plans for web use and a per-image API for automation, priced competitively.

The specific capability that matters: kerning. Letter-spacing. Ligatures that do not fall apart at small sizes. This is the thing every other model still gets subtly wrong when the prompt has more than ten words of visible type.

Where it falls short: weaker on photorealism, not a hero-image model for product mocks. Narrow tool, best in class for its narrow thing.

6. Adobe Firefly (brand-safe / enterprise)

What it does well: outputs you can ship in brand campaigns without a legal consult. Firefly is trained on Adobe Stock and openly licensed material, and Adobe indemnifies commercial use. That indemnification is the entire reason it lives on this list.

What it costs: bundled with Creative Cloud subscriptions. API access via Firefly Services is priced per-image for automation.

The specific capability that matters: enterprise-safe output. If you are producing anything that ends up in a paid ad, a brand kit, or a product packaging shot, Firefly is the safer default. No claim risk from training-data ambiguity.

Where it falls short: aesthetic ceiling is below Midjourney and below GPT Image 2 on product mocks. Firefly is the tool you reach for when legal blast radius matters more than maximum craft.

7. Recraft (vector + design systems)

What it does well: vector-native output. Icons, logos, product-design-system assets that need to drop into Figma as SVG, not as a bitmap. Photorealistic raster generation is fine too, but its moat is the vector side.

What it costs: free tier for exploration, per-image API pricing for production.

The specific capability that matters: SVG output that does not need a tracing pass. If your design-system needs a new icon family tomorrow, Recraft renders the set, you drop them in, you move on.

Where it falls short: not a hero-image model for landing pages. Narrow in the same sense Ideogram is narrow.

8. ByteDance Seedream (best non-English generalist)

What it does well: the model Chinese-market teams use for everything. Reads Chinese prompts cleanly, renders Chinese typography better than any other model on this list, and on English prompts sits somewhere between Nano Banana and FLUX on quality.

What it costs: per-image pricing via Volcengine or Doubao. Access is easier from inside mainland China than outside; Volcengine does support non-China accounts but expect account-setup friction.

The specific capability that matters: Chinese text rendering. If you ship content into a Chinese-language market, Seedream is the default. If you do not, it is a fine generalist, not the first choice.

Where it falls short: API access and region friction. Product mocks and UI scenes trail GPT Image 2.

Also on the field, worth tracking

xAI Grok's image side: fast, cheap, still behind on craft. Worth tracking as xAI closes the capability gap, not a daily driver today.

Tencent Hunyuan: open-weights Chinese model with a permissive license. A credible alternative if your legal posture needs open weights and FLUX does not fit.

Stable Diffusion: Stability AI's lineage continues. Worth watching for the next major release.

One-line recipe per job

• Hero images for landing pages: GPT Image 2 via ChatGPT Plus. Fallback: Nano Banana 2 on API.
• Product screenshots, UI mocks: GPT Image 2. Nothing else is close after today.
• Real product photography (cinematic, editorial, portraits): Midjourney.
• Logo-quality or dense-text design: Ideogram.
• Vector icons, design-system assets: Recraft.
• Brand-safe commercial ad creative: Adobe Firefly.
• Open-weights, self-host, legal clarity: FLUX.
• Bulk automation on Google: Nano Banana 2 (Flash tier).
• Chinese-market content: Seedream.

The shape of the field

A year ago, "image generation" was a two-horse race (OpenAI v1 versus Midjourney) with Stable Diffusion as the open-weights fallback and everything else a curiosity. Today it is a stack of specialists. The model ranking depends entirely on the job. There is no single "best image model" in April 2026. There is GPT Image 2 for product and UI work, Midjourney for aesthetic, Ideogram for type, Firefly for indemnification, Recraft for vectors, FLUX for self-host, Nano Banana 2 for cost, Seedream for Chinese, and a queue of contenders a month behind.

What today did was move GPT Image 2 from "catching up" to "the right default for the specific job most founders actually have", which is a real product screenshot that does not exist yet and needs to look like a real product. That is a different job than illustration, editorial, branding, or vector design.

The right move if you ship things is to pick one primary, pick one fallback, and stop benchmarking. I use GPT Image 2 primary, Nano Banana 2 fallback. I will revisit when the next serious FLUX release lands, when multi-image editing stabilizes, or when anyone ships a clean open alternative to GPT Image 2's UI-mock capability.

Pricing and exact version numbers move fast. The capability positioning above should be stable for a few months. Check current pricing and versions on each vendor's site before locking in a choice.

---

The launch-day build has the five-prompt comparison with scorecards and every output side by side. The landing page hero workflow is the end-to-end recipe for actually shipping a hero from prompt to deployed in under 20 minutes.