← sources

OpenAI launched GPT-6.2 Astra and GPT-6.2 Sol

source: reddit (reddit:OpenAI)published 2026-11-17 17:47ingested 2026-11-18 02:38

retained text

[the benchmarks](https://www.reddit.com/r/OpenAI/s/dQECDWdFiD)

**TL;DR:**

\- **GPT-6.2 Astra** scores 62 on AA Intelligence Index, 3 points below Fable 5.2 (65), GPT-6.2 Astra scores 174 on the Epoch Capabilities Index (ECI), on par with Fable 5.2.

\- Cost per task: 6.2 Astra: 4$, Fable 5.2: 26.54$.

**GPT-6.2 Astra** is 20% cheaper than 6 Astra—and 60% cheaper for cached inputs—at **$8 / $0.40 / $40** per 1M input, cached, and output tokens.

**GPT-6.2 Sol** stays at **$2 / $0.10 / $10**. Both new models use fewer output tokens.

\- **Benchmarks:** Astra 6.2 leads *GDPval v2* (+6.2 pts), *AA-3D* (#1, 1,318 Elo), *Terminal-Bench-Science 0.1* (+9 percentage points), and *Agents' Last Exam* (+5.2 pts), and more than doubles Claude Fable 5.2's pass rate on *Harvey LAB-AA v1.1* legal tasks. Sol nearly matches Astra 6 on *FrontierCode 2.0* at \~1/5th the cost.

\- **Safety & Availability:** Astra is OpenAI's first public model to reach **High** AI self-improvement risk (prompting an internal "Control Room" monitor), while both hit **Critical** cyber and **High** bio/chemical risk. Rolling out today to **Plus, Pro, Enterprise, and Edu** (ChatGPT Work, Codex, API), with **Ultrafast mode** expanding to Pro 200 starting with Sol.

document fetched 2026-11-18 00:28 · canonical https://reddit.com/r/OpenAI/comments/1xyr27j/openai_launched_gpt62_astra_and_gpt62_sol