On July 14, 2026, Moonshot AI released Kimi K3 — a 2.8-trillion-parameter open-weight MoE model with a 1M-token context window and native vision. Within days it produced the strongest benchmark results any open-weight model has posted against Claude: K3 beats Claude Fable 5 on 6 of 14 shared benchmarks, tops the Arena.ai Frontend Code Arena at #1 (1,679 points), and edges past Claude Opus 4.8 on the Artificial Analysis Intelligence Index (57 vs 56). Claude Fable 5 still wins the other 8 of 14 head-to-head benchmarks and leads that same index at 60.
| Benchmark | Kimi K3 | Claude Fable 5 | Margin |
|---|---|---|---|
| BrowseComp (web research) | 91.2 (state of the art) | 88.0 | +3.2 |
| SWE Marathon (long-horizon agentic coding) | 42.0 | 35.0 | +7.0 |
| Terminal-Bench 2.1 | 88.3 | 84.6 | +3.7 |
| Automation Bench | 30.8 | 29.1 | +1.7 |
| Program Bench | 77.8 | 76.8 | +1.0 |
| SpreadsheetBench 2 | 34.8 | 34.7 | +0.1 |
Pattern: K3's wins cluster in agentic execution — long-horizon coding runs, terminal work, browser research, and workflow automation.
| Benchmark | Kimi K3 | Claude Opus 4.8 |
|---|---|---|
| Artificial Analysis Intelligence Index v4.1 | 57 | 56 |
| Terminal-Bench 2.1 | 88.3 | 84.6 |
| SWE Marathon | 42.0 | 40.0 |
| Online Exp Bench | 75.5 | 65.9 |
| DECK-Bench | 73.5 | 66.9 |
| Finance-Bench | 62.6 | 58.4 |
| Benchmark | Claude Fable 5 | Kimi K3 | Margin |
|---|---|---|---|
| GDPval-AA v2 (professional work, Elo) | 1,760 | 1,668 | +92 — widest gap of any benchmark |
| FrontierSWE | 86.6 | 81.2 | +5.4 |
| JobBench | 57.4 | 52.9 | +4.5 |
| Kimi Code Bench 2.0 | 76.9 | 72.9 | +4.0 |
| Zerobench (visual, pass@5) | 46.0 | 41.0 | +5.0 |
| DeepSWE | 70.0 | 67.5 | +2.5 |
| CharXiv (visual, with tools) | 93.5 | 91.3 | +2.2 |
| AA-Briefcase (Elo) | 1,583 | 1,548 | +35 |
Overall standing on the Artificial Analysis Intelligence Index v4.1: Claude Fable 5: 60 · GPT-5.6 Sol: 59 · Kimi K3: 57 · Claude Opus 4.8: 56 · Kimi K2.6: 44. Moonshot's own launch materials place K3 third overall, behind Fable 5 and GPT-5.6 Sol — the wins above are per-benchmark, not an overall crown.
| Per 1M tokens | Kimi K3 | Claude Fable 5 |
|---|---|---|
| Input (cache miss) | $3.00 | $10.00 |
| Input (cache hit) | $0.30 | $1.00 |
| Output | $15.00 | $50.00 |
K3 is roughly 3.3x cheaper across the board at a 1M-token context. On BrowseComp, its state-of-the-art 91.2 comes in at under $2 per task, where competing frontier models run $5–$27 per task.