Concept Synth INDUCTION Challenge100 Round 1
General QA
Challenge100 combines the frozen Challenge64 deck with disjoint New36 tasks. Train accuracy counts direct Round-1 train-valid formulas over all 100 problems. Missing, provider-error, incomplete and parser-invalid responses receive no credit. Response cascades select by parser evaluability, never correctness or holdout outcomes. Multi-formula responses succeed if any formula is train-valid. This aggregate takes priority over its overlapping Challenge64 component in capability fitting.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 94.0 |
| GPT-6.1 Sol | 90.0 |
| Claude Opus 5.5 | 65.0 |
| GPT-6 Sol | 57.0 |
| Claude Sonnet 5.5 | 44.0 |
| GPT-5.6 Sol | 43.0 |
| Claude Fable 5.1 | 33.0 |
| Claude Fable 5 | 31.0 |
| GPT-5.6 Terra | 25.0 |
| Grok 4.6 | 25.0 |
| Claude Opus 5 | 24.0 |
| Grok 4.7 | 24.0 |
| Muse Spark 1.1 | 21.0 |
| GPT-6 Luna | 18.0 |
| GPT-5.6 Luna | 16.0 |
| DeepSeek V4 Pro 0813 | 15.0 |
| Gemini 3.7 Flash | 11.0 |
| Grok 4.5 | 11.0 |
| Muse Spark 1.2 | 11.0 |
| DeepSeek V4.1 Flash | 11.0 |
| Gemini 3.8 Flash | 9.0 |
| Gemini 3.5 Flash | 7.0 |
| DeepSeek-V4-Pro | 6.0 |
| Gemini 3.6 Flash | 5.0 |