Concept Synth INDUCTION Challenge64 Round 1
General QA · 2026-07-21
Challenge64 is a frozen 64-instance hard FullObs subset of INDUCTION. Train accuracy uses a fixed denominator of 64 and direct Round-1 responses before symbolic processing. Missing and parser-invalid responses count as incorrect; multi-formula responses receive credit if any submitted formula is train-valid. Holdout accuracy is a separate diagnostic. Challenge64 is a component fallback for the overlapping Challenge100 family.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 98.4 |
| GPT-6.1 Sol | 95.3 |
| Claude Opus 5.5 | 71.9 |
| GPT-6 Sol | 65.6 |
| Gemini 4 Argon | 65.6 |
| GPT-5.6 Sol | 57.8 |
| Claude Sonnet 5.5 | 50.0 |
| Claude Fable 5 | 42.2 |
| GPT-5.6 Terra | 37.5 |
| Grok 4.6 | 35.9 |
| Claude Fable 5.1 | 34.4 |
| Grok 4.7 | 34.4 |
| Claude Opus 5 | 32.8 |
| Muse Spark 1.1 | 32.8 |
| DeepSeek V4.1 Flash 1017 | 32.8 |
| GPT-6 Luna | 28.1 |
| GPT-5.6 Luna | 23.4 |
| DeepSeek V4 Pro 0813 | 23.4 |
| Grok 4 | 20.3 |
| GPT-5.4 | 18.8 |
| Gemini 3.7 Flash | 17.2 |
| Grok 4.5 | 17.2 |
| Muse Spark 1.2 | 17.2 |
| DeepSeek V4.1 Flash | 17.2 |
| Gemini 3.8 Flash | 14.1 |
| Mercury 3 | 14.1 |
| GPT-5.2 | 14.1 |
| Gemini 3.5 Flash | 10.9 |
| Kimi K3 | 10.9 |
| Grok 4.1 Fast | 9.4 |
| DeepSeek-V4-Pro | 9.4 |
| Opus 4.6 | 7.8 |
| Gemini 3.6 Flash | 7.8 |
| Kimi K2.7 Code | 7.8 |
| Opus 4.8 | 6.3 |
| Kimi K2.6 | 4.7 |
| Sonnet 5 | 4.7 |
| Gemini 3 Pro Preview | 4.7 |
| Grok 4.3 | 3.1 |
| Gemini 3.1 Pro Preview | 1.6 |