How close is AI to recursive self-improvement?
RSI (recursive self-improvement): in August 2026 Claude led 26% of Anthropic's own AI research work, up from under 1% in February 2026, and in October 2026 Claude Opus 5.5 scores 37 of 100 on Vals AI's test of doing AI research alone.
AI research score
37/100
Claude Opus 5.537
Vals' reference results50
Reaches 50
Q2 2027
if the measured pace holds
Slowest measured pace: Q2 2027 · Fastest: Q1 2027
AI is 13 points from Vals' reference results
| Date | Model | Score |
|---|---|---|
| Nov 2025 | Claude Opus 4.5 | 15 |
| Dec 2025 | GPT-5.2 | 16 |
| Apr 2026 | Claude Opus 4.7 | 19 |
| May 2026 | Claude Opus 4.8 | 24 |
| Jun 2026 | Claude Fable 5 | 24 |
| Jul 2026 | Claude Opus 5 | 33 |
| Sep 2026 | Claude Fable 5.1 | 36 |
| Sep 2026 | Claude Opus 5.5 | 37 |
Best now37Claude Opus 5.5, Oct 2026
First score15Claude Opus 4.5, Nov 2025
Pace18 to 28 points a year
Reaches 50Q2 2027
An AI agent trains AI models almost as well as their makers
An AI agent gets 10 hours and one AI chip to improve four small AI models. The makers trained their own versions with far more time and chips.
- Best so far
- Other models
| Date | Value | Model |
|---|---|---|
| 19 Feb 2026 | 18.1% | Gemini 3.1 Pro (OpenCode) |
| 16 Apr 2026 | 24.3% | Claude Opus 4.7 (Claude Code) |
| 23 Apr 2026 | 23.9% | GPT-5.5 xHigh |
| 28 May 2026 | 31% | Claude Opus 4.8 Max (Claude Code) |
| 9 Jun 2026 | 37.3% | Claude Fable 5 (Claude Code) |
| 9 Jul 2026 | 31% | GPT-5.6 Sol |
| 16 Jul 2026 | 31.5% | Kimi K3 |
| 24 Jul 2026 | 35.7% | Claude Opus 5 |
| 1 Sep 2026 | 44.7% | Claude Fable 5.1 (Claude Code) |
| 3 Sep 2026 | 41.9% | GPT-6 Astra |
| 22 Sep 2026 | 43.8% | Claude Opus 5.5 Max |
| 29 Sep 2026 | 38.8% | GPT-6.1 Sol |
Who still does the work
AI works with peopleOver 90%
AI leads, a person supervises26%
AI alone0%
Every model tested
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 5.5 | 37 |
| 2 | Claude Fable 5.1 | 36 |
| 3 | Claude Opus 5 | 33 |
| 4 | Gemini 4 Argon | 31 |
| 5 | GPT-6 Sol | 28 |
| 6 | GPT-6 Astra | 27 |
| 7 | GPT-6.1 Sol | 27 |
| 8 | Grok 4.6 | 25 |
| 9 | Wonder Full | 25 |
| 10 | Claude Fable 5 | 24 |
| 11 | Claude Opus 4.8 | 24 |
| 12 | GPT-5.6 Sol | 24 |
| 13 | GLM 5.3 | 23 |
| 14 | Qwen 3.8 Max | 22 |
| 15 | Kimi K3 | 21 |
| 16 | Grok 4.7 | 20 |
| 17 | Gemini 3.8 Flash | 20 |
| 18 | Muse Spark 1.3 Max | 20 |
| 19 | Claude Opus 4.7 | 19 |
| 20 | Gemini 3.7 Flash | 18 |
| 21 | GPT-5.5 | 16 |
| 22 | GPT-5.2 | 16 |
| 23 | GPT-5.4 | 15 |
| 24 | Claude Opus 4.5 | 15 |
| 25 | Kimi K2.5 Thinking | 7 |