TL;DR
OpenAI launched the GPT-5.6 model family (Soul, Terra, Luna), with the flagship Soul model having autonomously post-trained the smaller Luna model via Codex — a landmark recursive self-improvement milestone. The models outperform Claude Opus 4.8 and Claude Fable 5 on key benchmarks while being significantly cheaper, faster, and using fewer output tokens.
Key points
1
GPT-5.6 Soul autonomously post-trained Luna using a short Codex prompt, marking a real-world instance of AI-assisted AI training at frontier scale
2
GPT-5.6 Soul scored 53.6% on Agents Last Exam at a cost of $763, versus Claude Fable 5 at 40.5% costing $2,300 — better score at one-third the price
3
GPT-5.6 Soul achieved state-of-the-art 80 on the Artificial Analysis coding agent index, beating Fable 5 by 2.8 points while using less than half the output tokens at roughly one-third the cost
4
OpenAI launched ChatGPT Work, a new desktop app analogous to Claude Code, designed to connect tools, workflows, and documents for non-developers as well as engineers
5
A new Ultra mode extends reasoning beyond the previous max setting, and Cerebrus chip integration reportedly runs GPT-5.6 10x faster
Key takeaways
→
When evaluating frontier AI models, prioritize the cost-per-benchmark-point curve rather than raw scores alone — GPT-5.6 Soul currently leads this curve against Claude Fable 5 and Opus 4.8
→
Ultra mode in GPT-5.6 is worth testing for complex agentic tasks, but note that Max mode can slightly reduce benchmark scores, suggesting over-reasoning is a real risk to manage
→
For frontend design and computer use tasks specifically, GPT-5.6 Soul shows a meaningful qualitative leap — the creator recommends stress-testing these two capabilities together as a practical evaluation focus
Notable quotes
“It's pretty crazy that we've gotten far enough where you can have a fairly underspecified prompt and then just give it to Codex and then just have it go and run the job.”
“This isn't just a little tiny model release. There's nothing incremental about this. This is a beast of a release.”
“When the models reach a certain level in both those categories, kind of magical things begin to happen. And I think we are beginning to see it here.”
Worth watching?
✅
Worth watching the full video?
Watch if you want to see the live demos of GPT-5.6 building a zombie survival game and a sci-fi scale visualizer — the benchmark charts and cost comparisons are well covered here, but the visual output quality is best judged by seeing it yourself.
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