GPT-6.2 Astra tested: roughly on par with Fable 5.2, efficiency is the real story
kimmonismus · x · 2026-11-18
OpenAI released GPT-6.2 Astra and Sol, with Astra 20% cheaper than GPT-6.1 Astra. OpenAI claims it scores 5.5 points above Claude Fable 5.2 on Agents' Last Exam at 5% of the cost per task. However, the reviewer found it falls short of Fable 5.2 on benchmarks like Frontier Code 2.0, concluding the two are roughly on par rather than a clear win. GPT-6.2 Astra is also billed as the first public model to hit the "High" threshold for AI self-improvement. The reviewer notes OpenAI did not ship a model that truly hurts Anthropic two days before its IPO — an underwhelming release.
More from Models
- GPT-6.2 Astra finds vulnerabilities at ~$12 each, worth ~$3,000 in bounties — JiaweiLiu_ · 2026-11-18
- GPT-6.2 Astra scores 100% on ARC-AGI-3, first model to max the benchmark — patience_cave · 2026-11-18
- Leak suggests new Qwen4 models were not trained on cybersecurity data — mgostIH · 2026-11-18
- Teknium flags inconsistent OpenAI demos: was GPT-6 quietly nerfed before GPT-6.2? — Teknium · 2026-11-18
- GPT-6.2 Astra beats Opus 5.5 across agent benchmarks at 3-6x lower cost — VraserX · 2026-11-18
- Side-by-side: GPT-6.2 Astra builds noticeably richer scenes than GPT-6 Astra — Angaisb_ · 2026-11-18