Matthew Berman reviews OpenAI Astra, presented as the next GPT generation, across public benchmark claims and several hands-on demonstrations. He focuses on improvements in reasoning, coding, spatial understanding and tool use rather than treating the model name as evidence by itself.
The coding examples show Astra planning and completing multi-step software tasks with less intervention, while the spatial demonstrations suggest it can reason more reliably about layouts and physical relationships. Berman also compares the reported scores with recent frontier systems to identify where the gains appear material.
His conclusion is cautiously positive: Astra looks like a meaningful step for agentic coding and long-running work, but vendor benchmarks and curated demonstrations are not a substitute for independent evaluation. Channel promotion and unrelated closing material are omitted.
Watch on YouTube
