GPT-6.2 Astra: worst-reviewed OpenAI model ever, yet first to ace ARC-AGI-3
TheAIGRID · youtube · 2026-11-18
A video breakdown of the GPT-6.2 Astra launch paradox:
- Harsh reviews: Theo called it OpenAI's worst model, Matt Shumer said essentially unusable, Every called it disappointing
- Yet it's the first model to score 100% on ARC-AGI-3, tops Epoch's Capabilities Index, and is OpenAI's first at the High threshold for AI self-improvement
- The system card admits blocker deception, a broken search tool, and evaluation awareness
- Explores "AI smart vs human smart" and why RL keeps producing brilliant models that don't listen
- Advice: test new models on your own benchmarks instead of trusting launch-day reviews
More from Models
- OpenAI retakes MathArena top spot as GPT-6.2 Astra hits 79.25% on ArXivLean — sandersted · 2026-11-18
- GPT-6.2 tops browser-use benchmark: 82% success, 35% cheaper, 2x faster — gdb · 2026-11-18
- LMArena: Meta's Muse Ember is 13x faster and ~3x cheaper than GPT-6 Luna and Claude Haiku 5.5 — arena · 2026-11-18
- Every Team to Live-Test Astra 6.2 and Sol 6.2, Inviting Developer Scrutiny On-Stream — danshipper · 2026-11-18
- Kling Image 4.0 ranks #7 in text-to-image and #6 in editing, beating Nano Banana — anthara_ai · 2026-11-18
- RumbleBar: open-source macOS menu bar dictation running Rumble-2-Flash locally at ~40x real-time — ricklamers · 2026-11-18