GPT Image 2 vs the field

OpenAI shipped gpt-image-2 this morning. Before anyone had time to write a take, I ran the same five prompts through it, two older OpenAI models, and both of Google's Nano Banana variants, then saved every output and scored every axis.

Interactive grid above. Click any image to zoom. Toggle model columns on/off from the chips.

The headline

GPT Image 2 won every axis. 25/25.

Nano Banana Pro and Nano Banana 2 both scored 23/25. Still strong, but the text + UI lead Google had held for months flipped today.

The biggest gap was on the hardest prompt (three engineers at a whiteboard, handwritten labels, laptop with a Grafana dashboard). GPT Image 2 was the only model that drew all three engineers and a complete, internally consistent microservices diagram: Web Client → API Gateway → User / Order / Auth / Payment services → Postgres / Redis / Kafka. Every other model either missed an engineer, garbled a label, or drew a diagram that did not make engineering sense.

The five prompts and why each one

Designed to stress-test capabilities that matter if you build things:

1. Calibration, terminal + code. Pixel-perfect text rendering. Separates modern models from the DALL-E-era baseline instantly.
2. UI + 3 bullets. ChatGPT macOS window with a real-looking conversation. Tests the exact capability OpenAI was pitching today.
3. Brand hero. Dark landing page with headline, CTAs, a product mock with a score badge and a bar chart. Tests whether it can ship as real product imagery.
4. Portrait. Close-up photorealism, no text. Tests the old weakness.
5. Whiteboard. Three engineers, handwritten labels, laptop screen. Multi-subject scene consistency and in-scene text. The hardest prompt.

How I ran GPT Image 2 today

GPT Image 2 is behind an API gate (org verification required) at launch, so the cleanest way to run it today was the ChatGPT Plus web app with Thinking mode on. Google's Nano Banana Pro and Nano Banana 2 ran via Vertex AI. The older OpenAI models ran via the /v1/images/generations endpoint.

Surprises

The v1 → v1.5 jump was real. gpt-image-1 is DALL-E-3-era, with garbled text and flat product mocks. gpt-image-1.5 was what ChatGPT had quietly been serving pre-today, and it renders terminals and UI mocks cleanly. That is a story most people missed.

Nano Banana 2 ties Nano Banana Pro at half the cost. Same total (23/25). Pro is more literal, Nano Banana 2 adds atmosphere. If you are not using a Google-side model yet, start with Nano Banana 2.

Thinking mode on GPT Image 2 is doing work. It visibly planned layout before drawing (took 20-30 seconds to reason, then was fast on the image itself). That planning phase is probably why it got the microservices diagram right when no one else did.

The best landing-page hero came from v2. Prompt 02 asked for a hero for a build called "Landing Page Critic." GPT Image 2's output was more polished than most real product screenshots. Full nav, dark dashboard, B+ 8.4 hex badge, category scores, sub-metric cards, a recommendations panel. It is genuinely usable as a production hero.

Which model for which job

• Landing page heroes and product mocks: GPT Image 2. If API-only, fall back to Nano Banana 2.
• Portraits, photorealism: GPT Image 2 or Nano Banana 2. Nearly tied.
• Text-heavy UI mocks: GPT Image 2 > Nano Banana Pro > Nano Banana 2.
• Multi-subject editorial scenes: GPT Image 2 pulls decisively ahead. No close second.
• Bulk automation on Google: Nano Banana 2 (half the cost of Pro, same quality on most work).
• Do not reach for: gpt-image-1. Use 1.5 at minimum if you need an OpenAI fallback.

How this was built

• Google side: Vertex AI, gemini-3-pro-image-preview and gemini-3.1-flash-image-preview on my shipwithtez GCP project. Straight generateContent call with responseModalities: ["IMAGE"].
• OpenAI v1 and v1.5: /v1/images/generations with model: "gpt-image-1" and "gpt-image-1.5". No verification required.
• OpenAI v2: chatgpt.com with Thinking mode on. Saved the images out via DevTools. API access gated behind org verification (Stripe Identity).
• Scripts and dashboard: ~/MacminiDocs/MyProjects/images-shootout-apr21/, with run.sh for Google, run-openai.sh for OpenAI, and dashboard.html for the internal review. Runbook in 06-reference/learnings/2026-04-21-image-model-shootout-gpt-image-2-launch-day.md.

What I would do next

A follow-up workflow post is live: Prompt-to-production landing page hero with GPT Image 2. It covers going from a build brief to a shippable hero image in under 20 minutes, using the winning model from this comparison. That post swaps the Landing Page Critic hero on this site with the v2 output from prompt 02 above, as the before-after proof.