What's this about?
SpaceXAI released a larger model trained with longer RL on multi-hour tasks. It keeps 4.6 list pricing, adds native Grok Bot training and reports better coding, office-work and safety scores.
What exactly is long-horizon work?
It is a task with many dependent steps where an early error can corrupt everything later. A benchmark gain does not ensure the model will stay correct through a real repository or business workflow.
What's the community saying?
HN 555 points, 477 comments; official X post 26,500-plus likes at fetch time.
- Users report sharply different results across frontend, backend and agent harnesses.
- Readers question reversed axes, omitted comparisons and xHigh-versus-High settings.
- Cost per completed task and token efficiency matter more than list price alone.