RSI is here?! GPT 6.2 is “high” on recursive self-improvement

RSI is here?! GPT 6.2 is “high” on recursive self-improvement

More

Summary

David Ondrej covers the launch of OpenAI’s GPT-6.2 Astra and GPT-6.2 Soul, the first models OpenAI says rank high on its recursive self-improvement index and the first such class released publicly. He walks through the official announcement and the main benchmark charts to see how Astra compares with Claude Opus 5.5 and Fable 5.2.

Astra leads on several evaluations, including Agents’ Last Exam, Terminal Bench Science and a science index, and it does so at much lower cost per task. On the Artificial Analysis index Fable 5.2 still wins, reaching 64.6 at max effort versus Astra’s 60.2, but at roughly seven times the cost. Ondrej notes that extra-high effort sometimes outperformed max, and points out the competitive shift as Anthropic has overtaken OpenAI in revenue and valuation, while Gemini 4 Carbon gets a mention in the benchmarks.

He also surveys mixed reactions, from Matt Shumer calling the model essentially unusable to others praising it, then tests it himself in his Cloud Room agent interface. Tasks include UI animation design, bug fixing, and a 30-minute whole-codebase audit across 26 files, with direct comparisons to Fable 5.2. He also discusses subscription choices between Claude and OpenAI’s $200 and $500 plans.


📺 Source: David Ondrej · Published November 17, 2026
🏷️ Format: News Analysis

1 Item

Channels

2 Items

Companies