GPT-5.6 Soul Ran an Autonomous Ad Engine: 80+ Products via Codex + Higsfield MCP
Instead of benchmarking, Nick Saraev stress-tests GPT-5.6 Soul on a real-world task: he lets it run a fully autonomous "product ad engine." Driven from Codex CLI and wired to the Higsfield MCP (image + video models) plus Netlify for hosting, the model invents 80+ physical products, generates stills with GPT Image 2, animates ad videos, and auto-publishes sites — with a self-QA loop and parallel sub-agents. The core lesson: don't hand the model a one-to-one prompt; tell it to build the infrastructure to generate thousands and run itself, then let human taste filter the winners.
"I Gave GPT-5.6-Sol Unlimited Money to Make Ads (+ Results)" by Nick Saraev — Watch on YouTube →
Key Takeaways
- Let the model prompt itself. Rather than "make this one ad," Saraev tells GPT-5.6 Soul to build the infrastructure to generate millions of candidates and run on an autonomous loop with self-verification. Ideation at scale is what agents do far faster than humans; human taste then narrows the pile to a few winners.
- The stack. GPT-5.6 Soul (text/reasoning) via Codex CLI, GPT Image 2 for stills, the Higsfield MCP as a single gateway to many video models (Cling 3.0 Turbo, Seedance 2, etc.), and Netlify for instant auto-deploy of the generated sites.
- Ground video in images first. He generates high-quality stills before animating — feed video models raw and the output tends to be nonsensical.
- Self-QA loop + parallelism. Every asset passes a QA step: generate → the model critiques its own output with vision → regenerate. Sub-agents parallelize ideation → stills → video → publish, so a huge batch finishes in ~10 minutes instead of 5–10 minutes per product done linearly.
- Prompt phrasing that keeps it going. "You're running an unattended long-horizon creative session… The human is away. You will not receive answers to their questions. Do not stop to ask. Decide and proceed." This pushes the model to pick the highest-probability path and continue rather than block on you.
- Where it's heading. Saraev frames this as "reality shipping" — a pipeline that could scrape desired features, invent products, run and validate ads, then hand a spec to a machine shop. Expect a wave of fully AI-generated ads; verify that anything you buy is a real product.
Commands & Code Mentioned
# 1. Install and launch Codex CLI (copy the install command from the "get started
# with Codex CLI" page for macOS/Linux), then run:
codex
# 2. Pick the model and reasoning level
/model # select GPT-5.6 Soul, "medium" reasoning (quality vs. cost)
# (Saraev also enables a permissive "yellow"/YOLO-style mode — see caveats)
# 3. Connect the Higsfield MCP
# From Higsfield → "MCP & CLI" → ChatGPT/Codex → enable developer mode,
# then in Codex just ask it to set up Higsfield and approve the auth.
/mcp # list connected MCP servers + tools (confirm Higsfield loaded)
# 4. Give it a Netlify personal access token for auto-deploy
# Netlify → User settings → Applications → new access token → add to workspaceSaraev runs the agent in a permissive, low-confirmation mode so it won't stop to ask. He flags the risk himself: with bypass-permissions modes an unattended agent could go down a path you don't want (in his words, worst case it "deletes your entire hard drive"). Run this kind of loop in a sandbox/container, and note it consumes real GPT Image and Higsfield credits as it works.
More ChatGPT news
ChatGPT 5.6 + Codex Tested: Websites, Computer Use, Plugins & the Pet
GPT-5.6 Soul on Codex Ultra Autonomously Produced This Video
Run Codex CLI Free via OmniRoute: 93-Provider Gateway with Auto-Fallback
GPT-5.6 Government Freeze: From Model Wars to Context Wars (analysis, not a how-to)
ChatGPT Tutorial for Beginners: Get Real Work Done, Not Just Chat
GPT-5.5 Instant: OpenAI's faster, more accurate default model