Astra, First of the GPT-6 Family, Wows but Burns Billions of Tokens
Ben's Bites · rss · 2026-09-08
Ben's Bites reviews Astra, the first model in the GPT-6 family: the author burned 4B tokens over a weekend and calls it a token-hungry, spiky model — impressive but not consistently great, echoing the rough GPT-5 launch. Key points:
- Benchmarks & pricing: Strong on ARC-AGI-3, tops Zapier's AutomationBench, priced the same as Fable 5.1; community reviews are polarized.
- Standout demos: Drawing a portrait in Canva like a human, computer use fast enough to play piano, rebuilding Manhattan in Unreal Engine, autonomous UI generation, spectrogram-based sound identification, 3D-printed custom parts.
- Codex integration: Astra can ask questions without waiting, continuing non-blocking work without losing task context.
- Also: OpenAI hit its automated research intern goal (next: automated AI researcher by March 2028); Anthropic tests Claude Code plugins for behavior control (unshipped); Mistral raised €3B at €21B+ valuation led by Samsung; Inworld launched Realtime TTS-2 with sub-100ms P99 latency across 200+ languages.
More from coding & agent
- Denny Zhou: Least-to-Most Prompting Was an Early Agent Harness, Hitting 99% on SCAN — denny_zhou · 2026-09-08
- Google Cloud publishes developer guide to cutting Gemini token usage in agents — leslysandra · 2026-09-08
- Codex Windows app eats 30GB RAM at 6 tabs, users slam OpenAI over unfixed performance — karmay007 · 2026-09-08
- Mycelium report skill self-update tames mannered AI prose — arjunrajlab · 2026-09-08
- Spline's Astra agent and MCP let you build fully editable 3D scenes from prompts — dunkhippo33 · 2026-09-08
- Reddit debate: should research agents split search, verification, and synthesis? — mrclassical69 · 2026-09-08