Image Provider Shootout

One prompt, four generation routes. 2026-06-10. Merlino fleet routing test.

PROMPT: Create a beautiful, modern, professional infographic titled "MULTI-AGENT FLEET" showing one orchestrator agent node at top delegating with downward arrows to four specialist agent nodes labeled "RESEARCH", "DESIGN", "BUILD", "QA". Clean white background, flat vector style, render all text verbatim exactly as quoted. No watermarks, no extra text.

GPT Image 2, via codex-cli

Cleanest typography and layout discipline. Outlined icon nodes, perfect verbatim text. Rides the Codex subscription, zero marginal cost.

free with sub ~90s

ROUTE DOWN: login works, but ChatGPT's web UI changed and the proxy's image-detection selector no longer fires ("no image produced before timeout"). Needs a selector repair. GPT Image 2 still available via codex-cli (left).

GPT Image 2, via ChatGPT web proxy

Same model through the ChatGPT web UI driven by the local proxy on :8090. No API key needed. Currently broken at the UI-automation layer, not auth.

free with sub needs repair

Nano Banana Pro, via FAL

Richest visual treatment of the set: gradients, dimensional badge nodes, strongest designed feel. Current fleet default for NBP.

~$0.15/img ~30s

Nano Banana Pro, via Kie.ai

Flatter pastel take, title tucked into the orchestrator node. Verified fallback route when FAL errors or runs out of credit.

18 credits/img ~45s

Three-way side by side

GPT Image 2 vs FAL vs Kie.ai side by side