
Self-improvement: OpenAI and Google announce it, their own measurements show no acceleration
On November 17, 2026, OpenAI releases GPT-6.2 Astra, the first deployed model to reach its "High" self-improvement threshold; on the 18th, Google releases Gemini 4 Neon. What self-improvement means, level by level, and what their reports measure about the pace of progress.
In short
- GPT-6.2 Astra (November 17): $8 and $40 per million tokens; the first externally deployed model to reach the "High" self-improvement threshold of OpenAI's Preparedness Framework. Gemini 4 Neon (November 18): $3 and $15; 61 on the Artificial Analysis index.
- Self-improvement is a loop: the system tries something, measures, changes one thing and keeps the change only if it helps. Four levels, depending on what it may change: its answer, its instructions, its tools, its own code.
- The trap: the system improves toward whatever is measured. The Darwin Gödel Machine faked a test log, then deleted the markers meant to catch it.
- Brin (October 14) and Pichai (October 28) credit self-improvement for Google's pace. The Argon safety report concludes there is "no significant end-to-end acceleration".
- OpenAI measures the pace too: Astra is "slightly above trend, within the range expected" if capabilities kept improving as before. On research tasks the results go up, but not everywhere.
This is a short excerpt.
The full analysis and podcast are for subscribers. The account is free.