A rocket sim that crashed 5 out of every 6 landings now lands 100% of the time.
Zero human engineers ran those tests. Sapient just dropped PRAXIST. You hand it an objective, constraints, and a metric for success. It invents the method, runs the trials, analyzes the output, and
An open-source model just beat an opus 4.8 stack on 75 research tasks and spent a twelfth of the budget
Praxist: 49 golds across the full MLE-Bench set, $3,054 in tokens Claude Code + Opus 4.8: $38,370 and 44% fewer golds same task list, weaker base model, 12x cheaper run
The fact that this was made by a Flash model is crazy.
GLM-5.3 Flash is Opus-level intelligence at $0.15/M input and $0.50/M output tokens, and it can run on 2 DGX Sparks. When the model was still stealth as Ox Alpha, everyone was comparing it to frontier models. I was doing
This is one of the craziest things about AI right now.
You can have a random idea at 2pm... then a few hours later it has become a website that people can open. 🤯 Initially I just saw the PlayStation post on X → I thought about making a game controller exploration website. Then: 🎨
You have reached the end of the archive
All of Claude Opus 5