🔴LIVE: OpenClaw 4.20 OUT!! Is ChatGPT finally better than Opus???
TL;DR
OpenClaw 4.2.x did not make ChatGPT usable in Alex Finn’s main test — after a live upgrade and side-by-side build-off, OpenClaw running Opus 4.7 finished a playable 3D shooter while the new ChatGPT setup stalled, looped in planning, and ultimately said it would hand the job to Claude Code.
The killer demo was a 3D snowy city shooter prompt, and Opus actually shipped — Alex gave Henry (Opus 4.7), Jason (GPT-5.4 via ChatGPT OAuth), and later Hermes the same task: build a modern open-city FPS with buildings, cars, citizens, and snow; Opus/OpenClaw produced a working browser game with bullet marks, pedestrians, cars, and blood effects.
Alex’s core complaint about ChatGPT in agent workflows stayed the same: it ‘just made plans’ — he repeats that earlier ChatGPT-in-OpenClaw behavior was endless plan mode, and in this stream the GPT agent again spent roughly 30 minutes reasoning without writing code, then hit reconnect errors and gave up.
Hermes nearly stole the show by being easier to set up and more decisive — while OpenClaw fought through a painful multi-agent setup, Hermes spun up a fresh GPT-powered agent in one shot, prompting Alex to say he might switch platforms if Hermes lapped OpenClaw in the coding race.
Anthropic caught collateral fire for confusing product communication and possible compute pressure — Alex speculated Claude/Opus may feel worse because Anthropic is hungry for compute, pointing to Opus 4.6 getting ‘stupider,’ Claude Code changes on the $20 plan, and vague messaging around OpenClaw/CLI support.
The stream was also classic Alex Finn chaos: community hype, a $100,000 ASUS supercomputer, and a polarizing octopus logo reveal — he floated buying an ASUS GB300 Grace Blackwell machine with 750GB of VRAM, claimed two Vibe Coding Academy members sold apps for $250,000 each, and unveiled the new Henry Intelligent Machines branding that split chat 50/50.
The Breakdown
Live upgrade, big claim, and immediate skepticism
Alex opens by framing the whole stream around OpenClaw 4.20/4.2.x and one specific promise: ChatGPT is supposedly finally useful inside OpenClaw. He’s openly doubtful — in his words, ChatGPT in OpenClaw used to be “useless,” stuck forever in “plan mode,” while Claude would “drag its ass across the finish line” no matter what.
Community warm-up, women in AI, and the usual Alex Finn energy
Before the testing really starts, he spends a while hyping the chat, shouting out women in the community, and saying Fin Fam is the most female-friendly AI community on the planet. He mixes that with testimonials — one viewer says they went from $170/month to $5,000/month with Claude Code, and Alex says two Vibe Coding Academy students built apps and sold them for $250,000 each.
Creating the new GPT agent becomes a mess
The plan is simple: create a brand-new OpenClaw agent called Jason, wire it to ChatGPT OAuth, and compare it directly against Henry, his Opus-powered setup. In practice, the setup drags badly: OpenClaw keeps getting confused about Codex login, OAuth, Telegram bot setup, and status updates, and Alex gets increasingly annoyed that he has to keep asking what it’s doing.
Side quests: Henry branding and the $100K machine
In peak livestream fashion, Alex detours into a reveal of the Henry Intelligent Machines logo — mainly a pixelated octopus meant to symbolize a swarm of agents, with owl options as backups. Chat is brutally divided, which he actually seems to enjoy, then he pivots into polling viewers on whether he should buy an ASUS supercomputer for about $99,999 built around Nvidia’s GB300 Grace Blackwell superchip, saying it would have around 750GB of VRAM and should ship in July.
Alex’s bigger worldview: distribution, tutorials, and AI disruption
Between setup headaches, he drops a bunch of the broader thesis that powers his content. He says the future is “AI efficiency combined with distribution,” pushes people to build audiences on X and YouTube, and explains his mission as helping as many people as possible get through the disruption phase of AI and into the rewards phase.
The main event: same game prompt, three agent setups
Once Jason is finally alive, Alex gives the same prompt to three systems: OpenClaw on Opus 4.7, OpenClaw on GPT-5.4, and later Hermes on Opus 4.7. The task is intentionally ambitious: build a 3D first-person shooter in a modern snowy city with detailed buildings, weapons, cars, and civilians, and make it as high quality as possible.
ChatGPT stalls, Hermes starts strong, Opus recovers and wins
The GPT agent starts with a neat sounding plan, then gets stuck exactly how Alex feared: lots of reasoning, no code, reconnect errors, and eventually an admission that it should just hand the job to Claude Code instead. Hermes looks promising and is much cleaner operationally, but the actual winner is Henry on OpenClaw with Opus 4.7, which stops overthinking, says it’ll build the thing itself, and finishes first.
The playable result — and the final verdict
The winning game literally pops open on Alex’s screen unprompted, which delights him because that’s how he wants agents to behave. It’s dark and rough, but clearly working: snowy city vibe, moving cars, walking civilians, bullet marks, blood effects, and enough interactivity for Alex to laugh through a mini rampage — and with that, he effectively closes the case: Opus is still the champion, ChatGPT still isn’t there in OpenClaw, and Hermes remains the most credible challenger.