GPT-6.2 Astra tops Agents' Last Exam: +5.2 points at under one-third the per-task cost
OpenAIDevs · x · 2026-11-18
GPT-6.2 Astra sets a new high on Agents' Last Exam, scoring 5.2 points above GPT-6 Astra at less than one-third the per-task cost. The benchmark tests complex multistep professional work like campaign analysis, slide creation, and financial spreadsheets. The quoted tweet adds: from public photos and build manuals of an RC buggy alone, Astra produced an animated assembly guide and an illustrated parts catalog; on BenchCAD 2.1 Core (building 3D parts from engineering drawings) it performed 32% better than GPT-6 Astra.
Related event: GPT-6.2 Astra Tops Agents' Last Exam Benchmark(2 posts)→
More from coding & agent
- tldraw office hands-on: sketch specs on a whiteboard together with agents — max__drake · 2026-11-18
- OpenAI DX engineer's article: turning GPT-6.2 Astra into a video editor with Codex — nickbaumann_ · 2026-11-18
- OpenAI says start daily coding with GPT-6.2 Sol: near-flagship scores at ~1/5 the cost — OpenAIDevs · 2026-11-18
- LangSmith adds trace access controls so engineers debug agents without exposing sensitive data — LangChain · 2026-11-18
- LangSmith adds three controls for sensitive trace data access — LangChain · 2026-11-18
- Tejas Kumar's Daiku agent harness filled out his Shark Tank application—and got him on the show — TejasKumar_ · 2026-11-18