Codex Just Made Claude "Obsolete"? — 9 Platform Moves Worth Studying

TL;DR

A creator ran through nine things Codex can do that Claude Code (mostly) can't yet — and the interesting thread isn't "Claude is obsolete" (it isn't), it's that OpenAI is winning on platform, not reasoning. The standouts for anyone building agent workflows: Codex auto-discovers the right skill from your prompt without you installing it; every paid ChatGPT user now gets a free cloud VM (≈15 GB RAM, 9 cores) that is the agent's own computer — it keeps working when your laptop is closed, and someone had it download and run a local Gemma model inside itself in 14 minutes; and ChatGPT Sites turns "build me an app" into a live, hosted product with a database, file storage, a secrets vault, and ChatGPT-account auth — all from a single conversation. The rest (digital pets as a status indicator, ChatGPT Work as an ambient phone-first agent, Codex + Remotion/Slides, a free security-audit plugin, plain-English PR summaries) rounds out the picture. My take: these are distribution and infrastructure advantages, not proof of a smarter model — so the honest posture is still both tools, but the platform gap is the thing to watch.

Key takeaways

• Skills that route themselves. Codex read a prompt and autonomously pulled the right skill (agent-reach — the repo that briefly hit #1 trending, 60k★) without it being installed. That's progressive disclosure at the platform level — the agent finds the tool, you don't wire it up.
• The "free computer" is the real story. A cloud VM per paid user (≈15 GB RAM / 9 cores; +5 GB on Pro) that runs independently of your machine. Renting the equivalent is ~$20–50/mo; here it's bundled — and it's the substrate everything else runs on.
• ChatGPT Sites ≠ website builder. Describe an app → it ships with a real DB, file storage, a vault for API keys/secrets, and auth via ChatGPT accounts (it handles the OAuth you'd normally dread), hosted, with custom domains. An a16z partner had it build + host + wire-up a GeoGuessr-style AI-vs-AI game inside one chat.
• Ambient, phone-first agents (ChatGPT Work). People are handing it standing jobs — watch an inbox, flag weird finances, hunt Airbnbs with its own remote browser, unsubscribe a parent from spam — run from a phone, no laptop needed. The cloud VM is what makes "switch it on and forget it" possible.
• Image-gen as an unfair advantage. Pair Codex with Remotion (video-from-code) or Codex Slides (repo/PDF → animated deck) and the best-in-class image generation becomes a media pipeline Claude can't match without bolting on a third-party key.
• Free security-audit plugin. Point it at a codebase → threat model, attack paths, fixes + tests, exported to GitHub/Jira/Linear or SARIF. Genuinely useful if you ship vibe-coded apps to real users.
• Plain-English PR summaries. Non-technical builders get a readable summary of what the agent actually changed — small feature, big trust unlock.
• The honest caveat: "obsolete" is clickbait. None of this is about the model being smarter — it's platform, distribution, and infrastructure. Claude competes on other axes. Study the moves, keep both tools.

---

The map: 9 moves, one strategy

Nine features, but they ladder up to a single bet — own the whole stack around the model so the agent has somewhere to live and something to ship to.

flowchart TD
    VM["☁️ Free cloud VM per user<br/>≈15 GB RAM · 9 cores · runs when laptop is closed"] --> SUB["The substrate everything runs on"]
    SUB --> SK["Skills auto-discovery<br/>(finds agent-reach uninstalled)"]
    SUB --> SITES["ChatGPT Sites<br/>host + DB + storage + secrets + auth"]
    SUB --> WORK["ChatGPT Work<br/>ambient, phone-first standing jobs"]
    SUB --> MEDIA["Remotion / Slides<br/>image-gen → video & decks"]
    SUB --> SEC["Security-audit plugin<br/>threat model → fixes → SARIF"]
    SITES --> DIST["Distribution:<br/>note-on-your-phone → live product in one chat"]
    style VM fill:#dbeafe,stroke:#2563eb
    style DIST fill:#dcfce7,stroke:#16a34a

Skills that find themselves

The sharpest builder-relevant moment: Codex looked at a prompt, decided it needed a skill it didn't have installed, and pulled in agent-reach (the skill that gives agents access to X/Reddit without paid API keys, and auto-switches to a backup when a method gets patched). The point isn't the one skill — it's that the platform does the routing. That's the same progressive-disclosure principle I want in my own setup, moved from "my CLAUDE.md routes to files" up to "the agent selects the capability."

Pets as a status indicator (the quirky-but-clever one)

Codex has Tamagotchi-style digital pets — nine to pick from, or generate your own. It's easy to dismiss, but the actual function is a glanceable status light: pet sleeping = agent idle, pet busy = agent cooking. It's a community/personality play (Claude, by contrast, "has a spinner"), and a reminder that adoption is partly emotional.

ChatGPT Sites — the move I care about most

This is the one that matters for building. When Codex "builds you a site," you actually get a real database (your app remembers users and their data), file storage, a vault for secret keys (API keys/passwords, not "taped to the fridge"), and auth through ChatGPT accounts — it handles the OAuth, the scariest part. Then it hosts it, with custom-domain support.

The distance from "$1M app idea rotting in a phone note" to "live product with real users" used to be months and thousands of dollars of Cloudflare-to-Stripe-to-Supabase plumbing. The demo that made the point: an a16z partner had it build, host, and wire up login for a GeoGuessr-style game where you play against AI models — inside a single conversation.

Why it's on my radar: this is OpenAI's answer to Claude Artifacts — but Artifacts don't ship hosting + access control + a DB + a secrets vault. For CKA that's internal dashboards/trackers deployed in minutes; the catch is no live-data connection yet, so you keep it fresh with a scheduled automation — which maps cleanly onto a Hermes cron feeding a site. (I kept a separate private note on ChatGPT Sites from OpenAI's own docs.)

The free computer (what everything runs on)

Every paid ChatGPT user now gets a full cloud VM — a whole computer in a data center, ≈15 GB RAM and 9 cores (Pro adds ~5 GB), with its own browser that can install whatever packages it needs. Because it's not your machine, the agent keeps working when your Wi-Fi drops or your laptop closes. The flex that sold it: someone challenged Codex to run a local model with the network restricted — it built its own network proxy, downloaded Gemma, and had it running in 14 minutes. An AI ran another AI inside itself.

What I'm taking from it

flowchart LR
    A["Platform > model IQ<br/>the gap here is distribution/infra,<br/>not reasoning"] --> T["Keep both tools;<br/>watch the platform moat"]
    B["Agent needs its own computer<br/>(cloud VM = ambient, always-on)"] --> H["Design Hermes for<br/>always-on cloud execution"]
    C["Build + host + DB + auth<br/>in one conversation"] --> D["Ship internal CKA dashboards<br/>this way; refresh via cron"]
    E["Skill auto-discovery"] --> F["Route capabilities,<br/>don't hand-wire them"]

• "Obsolete" is a thumbnail, not a finding. These are platform/distribution/infra wins. The right read is the moat is moving to the platform layer — so evaluate agents on what they can host, remember, and ship, not just how they reason.
• The cloud VM is the primitive to internalize. "The agent has its own always-on computer" changes what ambient automation can look like — it's the shape I want for Hermes.
• ChatGPT Sites is a real option for CKA internal tools — build+host+DB+auth in one chat, kept fresh by a scheduled automation. First candidate: a bookings/ops dashboard.
• Skill auto-discovery validates the routing thesis — progressive disclosure, but the platform selects the capability. That's the direction to push my own skills toward.

---

Synthesis and diagrams are mine; screenshots are from the creator's screen shares for study/commentary. This was a fast, sarcastic hype tour — I've kept the useful signal and dropped the "Claude bad" bit.