GPT-6 Astra hands-on: solves Blender, trained on 100k GPUs, compute still the bottleneck
thedealdirector · x · 2026-09-08
A hands-on reviewer shares takeaways from extensive testing of GPT-6 Astra:
- Accelerating capability: model performance keeps accelerating; GPT-6 has effectively "solved Blender" for most use cases, and the author jokes the next run should kill SAP Concur login chores.
- Compute is the constraint: GPT-6 was trained on 100k GPUs, and the trade-off between investing compute in training runs and serving inference remains the primary limiting factor—user interaction feedback is a critical part of training data.
- Transparency erosion: the author argues GPT-6 is a very close step toward AGI but will push other labs to obscure their chain-of-thought observability.
More from Models
- GPT-6 Astra Beats Atari's Montezuma's Revenge in Real Time With a Basic Harness — swishfever · 2026-09-08
- Claim: GPT-6 Astra beats Atari's Montezuma's Revenge in real time with a basic harness — swishfever · 2026-09-08
- Reddit Guide: Taming Qwen3.8 Overthinking with Task-Scaled Reasoning Budgets — Express_Quail_1493 · 2026-09-08
- GPT-6 Astra Taught Itself Factorio in 15 Minutes, Building a Working Mine — Cklly2004 · 2026-09-08
- ChatGPT Pro Page Lists Unlimited Access to 5.6, Sol, Terra, Luna Under Fair Use — Invalid-Function · 2026-09-08
- Gary Marcus says LLMs still haven't produced new linguistic generalizations, echoing Chomsky — GaryMarcus · 2026-09-08