TL;DR
Matthew Berman covers the release of GPT-5.6 (codenamed 'Soul') and demos impressive long-running agentic builds including a Minecraft clone, Excel clone, and Rubik's cube simulator created entirely by looping GPT-5.6 in Codex for up to 6 days. He also introduces two open-source tools his team built: a GPT-5.6 relay skill and 'Loopy', a loop design assistant.
Key points
1
GPT-5.6 ('Soul') is the final model of the GPT-5 pre-training run and is priced at $5/million input and $30/million output — identical to GPT-5.5 — making GPT-5.5 effectively obsolete overnight.
2
Soul uses far fewer tokens per task than Claude Fable, making its real cost-per-task significantly lower despite similar per-token pricing; Berman observed it takes the most direct path to a solution of any model he has used.
3
Berman ran 6-day unattended Codex loops to build a Minecraft clone with biomes and NPCs, a fully functional Excel clone (built by having Codex use computer-use to click around the real Excel), and a Rube Goldberg physics game.
4
Codex browser control was highlighted as a standout feature: Berman used it to automate DNS migrations across Vercel, DigitalOcean, and GoDaddy, and to auto-scale a Supabase instance during a large data import.
5
Two open-source tools were announced: a GPT-5.6 relay skill (github.com/forwardfuture) that orchestrates Soul for planning and delegates to Terra/Luna for coding, and 'Loopy', a loop design and monitoring skill with 2,500+ GitHub stars.
Actionable insights
→
Stop using GPT-5.5 immediately — Soul is the same price, meaningfully better at long-horizon tasks, and should replace it for all use cases.
→
Use Soul for planning and spec-writing, then delegate actual code writing to Terra or Luna to achieve roughly 50% cost reduction while maintaining similar output quality.
→
Install Loopy in Codex, Cursor, or Claude Code to design, monitor, and improve agentic loops — loops are conceptually simple but design quality varies widely and Loopy helps bridge that gap.
→
Enable Codex browser control for repetitive web tasks like DNS management or cloud resource scaling — once cookies are stored, it runs autonomously with minimal prompting.
→
When building with GPT-5.6, expect a default design aesthetic resembling PowerPoint slides; steer it with a few iterations to get significantly better visual output.
Notable quotes
“GPT-5.6 sees things directly. It is like, I am going to get this thing done. I know exactly how to get there. I am not going to straggle. I am not going to go off course. I am just going to get this thing done that you asked for and nothing else.”
“They are the same exact price. I assume they use the same exact quota. And so there really is no point in using 5.5 anymore. Crazy.”
“I gave it a goal. I said slash-goal create an Excel clone full feature parity. That was the entirety of the prompt I gave it.”
Worth watching?
⏭️
Worth watching the full video?
Skip the video — the key pricing data, capability comparisons, and demo descriptions are all captured here; the live footage of the Minecraft/Excel builds adds some visual interest but not enough to justify the unstructured livestream format.
Saved you some time? The creator still deserves a like.
Found this by searching? That's the hard way.
summree watches the AI channels you follow and sends you the summary the moment a video drops. Nothing gets past you. From £4/month.
Try it free