Vibe coding

One month with Codex on ChatGPT Plus (and what I built for OpenAI Ads)

OpenAI put a free month of Plus on my account. I used it for real Codex work — and shipped two tools for ChatGPT Ads while ads rolled into my region.

~11 min read

OpenAI put a free month of ChatGPT Plus on my account. I used it while it was there — mostly inside Codex for real work, not demo scrolling.

Some takes call Codex hype. After a month: it is not. Usage windows and slower completion vs my Cursor + Grok loop annoyed me. Design and architecture? Genuinely strong. Clear answers. Good questions. That part is not marketing fluff.

This is a warm field report: how I installed it, how I routed Sol / Terra / Luna, what I built for OpenAI Ads once ads showed up in Hungary and nearby markets, and where I still open Cursor first.

What Plus actually unlocks for Codex

ChatGPT Plus is $20/month. For Codex specifically, the current public map includes:

• Codex on the web, CLI, IDE extension, and iOS
• The GPT-5.6 family in Codex: Sol, Terra, Luna
• Cloud-side extras (code review / Slack-style integrations where available)
• Credits to keep going after included limits

Pro starts higher (from $100/month) with 5× or 20× more Codex usage than Plus. I tested on Plus, not Pro.

OpenAI’s own framing (paraphrased from the pricing page):

Back when GPT-5.6 Sol / Terra / Luna first shipped, access was gated. That chapter is over for paid ChatGPT users — the interesting question now is how you spend the allowance, not whether you can log in.

Every way to run Codex

I lived in VS Code + Codex. You do not have to — pick the surface that matches how you actually work.

IDE extension

Official path: Codex IDE docs

1. Install / enable Codex for your editor (marketplace download links are on that page — works with VS Code forks like Cursor and Windsurf).
2. Open the Codex sidebar (or Command Palette → Codex: Open Codex Sidebar).
3. Sign in with ChatGPT (or an API key).
4. Keep quick edits local; hand longer jobs to cloud/web when the task outgrows the sidebar.

I chose VS Code + Codex so Cursor stayed clean as my primary IDE. You can also run Codex inside Cursor via the same extension family — I preferred separation.

CLI

Official path: Codex CLI

curl -fsSL https://chatgpt.com/codex/install.sh | sh

Then open a project folder and run codex. First launch: Sign in with ChatGPT.

Handy in-session: /model, /status, /permissions, /init (scaffolds AGENTS.md). Also documented: npm global @openai/codex, Homebrew cask, Windows PowerShell installer — see the CLI page.

Desktop app

Official path: desktop app

1. Install ChatGPT for Mac / Windows / Linux.
2. Sign in.
3. Choose Codex (or ChatGPT, then Chat vs Work).
4. Start New chat and point it at a folder when you want local files in play.

This is also the on-ramp for computer use, scheduled tasks, and (where available) record-and-replay style workflows.

Mobile remote

Official path: Remote / connections

Codex does not magically run a full agent only on your phone. The pattern is: desktop host stays online, phone steers.

1. On the Mac/PC: ChatGPT desktop → Settings → Connections → Control this Mac or PC → set up.
2. Pair with the phone (same ChatGPT account).
3. From mobile Remote, pick the computer, continue the task, approve commands, review diffs.

Plus includes Codex on iOS in the current pricing list — remote control is how you stay in the loop on the tram without abandoning a long run.

Web Codex

Sign in at ChatGPT / Codex web for cloud tasks and review flows. Local vs cloud share the same agentic allowance; cloud chats often lean on Sol and can burn more of the window than a tight local edit (pricing notes).

My routing: Sol → Terra → Luna

This was my default for the month:

A coaching loop that matched reality:

1. Sol once for a hard plan (no code yet).
2. Terra for most of the build.
3. Luna for the laundry list.
4. Sol again for a milestone review.

That is close to what I did — I just over-used Sol early because it felt smarter. It is. It is also the fastest way to empty the tank.

Chat vs Work

I brainstormed and designed in Chat. Some people swear by the Work tab for merging spec files into one run. I did not love Work for my style — it felt like an extra surface when an extended Chat thread already had the context.

Dear readers: what are you doing? Chat-native, Work-tab, pure CLI, desktop-only? Tell me. I will steal the good patterns.

Threads beat one giant chat

Advice that matched my month: fresh threads for separate features. A clean context outperforms a 120k-token archaeology dig. Spin parallel threads when the work does not collide; summarize back into one note when you need the map.

Self-scoring loops (define success → agent scores itself → fix until it passes) are useful when you have a clear bar — tests green, computer-use walkthrough, or a reference set. Less magic when “good design” is subjective.

What I built for OpenAI Ads

OpenAI Ads expanded across Europe (Hungary included in Ads Manager availability). I am not an agency. I still wanted something any merchant could run before they paste a feed into Ads Manager.

So I built two Actors (same craft path as how I create Apify Actors people actually pay to use). Codex helped with design and iteration; the Store is how I shipped. This is proof of the month — not the whole story.

OpenAI Ads Readiness Checker

apify.com/kondasviktor/openai-ads-readiness-checker

Evidence scan: HTTPS, robots.txt for OAI-AdsBot (required) and optional OAI-SearchBot, landing/image reachability, optional Merchant feed sample. Independent tool — not affiliated with OpenAI. A green report is not an approval letter.

PPE from the Store page: about $0.15 per domain readiness report (with cheap extra URL checks).

Google Merchant → OpenAI Ads CSV

apify.com/kondasviktor/openai-ads-feed-converter

Takes Google Merchant XML/CSV/TSV/JSON and emits OpenAI Ads canonical CSV with row validation. Verified product-feed markets in the current pin: US / CA / MX. Other countries (including HU) are marked not-yet-verified — we do not fake readiness. v1 downloads CSV; you upload in Ads Manager yourself.

Rough PPE: $0.50 / 1,000 processed product rows.

New to Apify? Create an account (referral link — same as on our Tools page).

Honest cliffs

Read this before you cancel Cursor.

1) Five-hour windows still exist

You will see takes that “everything is only token credits now.” The practical Plus experience is still: five-hour window + weekly limits, then credits if you want to keep going (usage / pricing docs). Model choice and task size change how fast the window drains. Hitting a wall mid-prompt is the worst version of this — worse than Cursor’s usual flow for me.

Check /status in the CLI. Check Settings → Usage before a big job.

2) Slower than my Cursor + Grok 4.6 loop

Same class of task often took longer in Codex. Longer runs → more tokens → less allowance left. Quality on design/architecture still impressed me. Throughput is where Cursor won.

3) Long unsupervised runs burn the window

Long unsupervised Terra runs burned my Plus window fast while the job was still only partly done. Great intentions. Bad burn rate for “just execute the checklist.”

Takeaway: Codex is excellent as an architect and senior pair. Treat long unsupervised execution like a budget decision, not a default.

4) Stretch the allowance

Official tips that matched reality:

• Smaller, precise prompts; less junk context
• Lean AGENTS.md; fewer MCP servers when idle
• Terra/Luna for routine work
• New thread > one endless chat
• Credits or a reset if you are eligible and mid-deadline

So… is Codex hype?

No. Sol’s reasoning is the real deal for vibe coders who can write a clear brief. Terra is the daily driver. Luna is how you stop lighting money on fire.

Also no: it does not replace my primary IDE. Cursor is still home. Codex is the Plus-powered second brain I open when I want a strong design pass, a CLI agent, or mobile remote on a long job — and when I am shipping something like OpenAI Ads tooling.

If your primary is already Claude Code, Muse, or pure Cursor agents, that is fine. Different harnesses, different walls.

What do you prefer as your primary IDE right now? Reply and tell me — Cursor, Codex, Claude Code, or something weird. I will collect the patterns.

Quick start checklist

1. Decide surface: IDE sidebar, CLI, or desktop.
2. Sign in with Plus (or Pro).
3. /init or write a short AGENTS.md.
4. Sol once for the hard plan → Terra for the build → Luna for cleanup → Sol for review.
5. Watch Usage before unsupervised marathon jobs.
6. If you are staring at ChatGPT Ads: run the readiness checker, then the feed converter if you already have a Merchant feed.