| Kimi K3 | Claude Fable 5 | |
|---|---|---|
| arenaElo | 1478 | 1492 |
| humanEval | 96.5 | 98 |
| sweBench | 75.5 | 77 |
| mmlu | 94 | 95.1 |
| gpqa | 87.5 | 92 |
| aime | 92 | 94 |
| Price in/out ($/M) | $3 / $15 | $10 / $50 |
Fable 5 keeps the overall capability crown (Artificial Analysis 60 vs 57) and stronger long-horizon agent reliability. K3 counters with a third of the price ($3/$15 vs $10/$50), open weights due July 27 under a modified MIT license, and wins on some published coding benchmarks. For the hardest agentic and research work, Fable 5. For cost-sensitive volume, self-hosting or data-residency requirements, K3 is the strongest open option ever, but verify Moonshot's self-reported numbers on your own workload.