Comparing 2 models
Gemini 3.8 Flash vs Kimi K3
Benchmarks through
Verdict
Gemini 3.8 Flash places higher on 5 of the 5 tasks where both are ranked. Kimi K3 costs 3.5x less per input token. Only Kimi K3 has open weights for self-hosting.
- Choose Gemini 3.8 Flash for Fast Summaries (#1), Value LLM (#2), Massive Context (#4).
- Choose Kimi K3 for you need to self-host.
Built from the published task rankings below; ranks come from each task's own weighted benchmarks, so a win on one task says nothing about another.
Measured head to head
Benchmarks both models were run on
| Evaluation | Gemini 3.8 Flash | Kimi K3 | Source |
|---|---|---|---|
| Design Arena Coding | 1304 Elo | 1361 Elo | source · 2026-10-21 |
| Design Arena Full-stack Agent | 1244 Elo | 1315 Elo | source · 2026-10-14 |
| Design Arena UI | 1318 Elo | 1346 Elo | source · 2026-10-21 |
| Design Arena Web-app Agent | 1237 Elo | 1296 Elo | source · 2026-10-21 |
| Design Arena Website | 1301 Elo | 1335 Elo | source · 2026-10-21 |
| FrontierMath Tiers 1-3 (Epoch AI) | 68.4% ± 2.8 | 72.2% ± 2.7 | source · 2026-09-02 |
| GPQA Diamond (Epoch AI) | 95.4% ± 1.4 | 93.1% ± 1.5 | source · 2026-09-02 |
| LiveBench Agentic Coding | 54.2% | 62.2% | source · 2026-10-18 |
| LiveBench Coding | 72.5% | 81.5% | source · 2026-10-18 |
| LiveBench Data Analysis | 54.0% | 78.7% | source · 2026-10-18 |
| LiveBench Instruction Following | 81.4% | 71.4% | source · 2026-10-18 |
| LiveBench Language | 87.8% | 85.5% | source · 2026-10-18 |
| LiveBench Math | 91.6% | 84.4% | source · 2026-10-18 |
| LiveBench Reasoning | 89.3% | 90.7% | source · 2026-10-18 |
| LMArena Agent | 3.6 | 1.8 | source · 2026-10-20 |
| LMArena Agent Bash Recovery | -0.7 | 4.4 | source · 2026-10-08 |
| LMArena Agent Steerability | 1.3 | 0.4 | source · 2026-10-20 |
| LMArena Agent Task Outcome | 3.9 | 6.3 | source · 2026-10-20 |
| LMArena Agent Tool Hallucination | 0.6 | 0.7 | source · 2026-10-20 |
| LMArena Creative Writing | 1488 Elo | 1457 Elo | source · 2026-10-20 |
| LMArena Instruction Following | 1492 Elo | 1485 Elo | source · 2026-10-20 |
| LMArena Long Query | 1507 Elo | 1494 Elo | source · 2026-10-20 |
| LMArena Math | 1532 Elo | 1494 Elo | source · 2026-10-20 |
| LMArena Text | 1496 Elo | 1488 Elo | source · 2026-10-20 |
| LMArena Text Factuality | 1473 Elo | 1472 Elo | source · 2026-10-20 |
| LMArena WebDev | 1585 Elo | 1655 Elo | source · 2026-10-19 |
| LMArena WebDev Image-to-Web | 1593 Elo | 1579 Elo | source · 2026-10-19 |
| OTIS Mock AIME (Epoch AI) | 98.9% ± 0.9 | 97.2% ± 1.1 | source · 2026-09-02 |
| SimpleQA Verified (Epoch AI) | 69.7% ± 1.5 | 50.6% ± 1.6 | source · 2026-09-02 |
Values are as published by each evaluator, in its own units; rows from different evaluators are not on a common scale.
task evidence
How they place when the question changes
Lower ranks are better. “Best” marks the strongest option in this shortlist for that task.