MatthewBerman's early GPT-6 Astra review: best model yet, stunning browser control
MatthewBerman · x · 2026-09-04
Reviewer MatthewBerman, with early access, tested GPT-6 Astra extensively and calls it the best model he's ever used — a massive leap rather than an incremental upgrade, especially on zero-shot prompts.
Key takeaways:
- Knowledge work is the core: slide creation, analysis, writing, browser control; browser control is another level above GPT-5.6 and significantly faster
- Prompt-to-playable game has essentially arrived: the games are actually fun; he believes someone with a good eye could build a viral game in 1-2 weeks with Astra
- Writing: removes much of the "AI smell" though some remains; more steerable in design but tends toward GPT-5.6's forest-green aesthetic
- Highly steerable: adding prompt specifics greatly extends how long it keeps working (30-minute tasks)
- Unmatched 3D understanding: consistent 3D asset creation and impressive spatial awareness in complex world-building
He's making Astra his go-to model for difficult work.
Related event: Matthew Berman's GPT-6 Astra Review: Strongest Model Yet(4 posts)→
More from Models
- GPT-6 Astra hailed SOTA at 3D modeling with wild design output — dejavucoder · 2026-09-04
- GPT-6 Astra system card: first OpenAI model to hit Critical cybersecurity capability — mallow610 · 2026-09-04
- Astra keeps better theory-of-mind, cutting drift on long tasks — emollick · 2026-09-04
- Fchollet: ARC-4 due early 2027, final ARC likely 6-7, AGI around 2030 — MartinSignoux · 2026-09-04
- GPT-6 Astra claims first improvement on prime gaps since the 1930s, shaving a log log n factor — rapha_gl · 2026-09-04
- Leaked GPT-6 'Astra' reportedly builds Unity 3D cities, 1.9x faster than GPT-5.6 — Dr_Singularity · 2026-09-04