Harness and model combinations: which one is the best?
retained text
It's hard to compare harness x models for real life usecase, so I created a benchmark for this. It test basic math, vision, computer use (reading emails, navigating stores), and coding. It compares various models and various harnesses and see how good/bad and, more importantly, how fast they are. Top 10 by capability 1. Claude Code/Opus 5.5 - 100% (14m 26s) 2. openclaw/openrouter/qwen3.8-max-0902 - 98% (26m 33s) 3. opencode/openrouter/deepseek-v4.1-flash - 96% (10m 56s) 4. hermes/rtx5090/qwen3.8-27b-nvfp4 - 96% (1h 50m) 5. Codex Sol 5.6 Medium - 96% (21m 54s) 6. hermes/openrouter/qwen3.8-max-0902 - 96% (1h 23m) 7. pi/openrouter/deepseek-v4.1-flash - 94% (13m 36s) 8. opencode/rtx5090/qwen3.8-27b-nvfp4 - 94% (1h 41m) 9. opencode/openrouter/glm-5.3-flash - 92% (22m 54s) 10. pi/openrouter/glm-5.3-flash - 92% (34m 04s) Top 10 by speed 1. openclaw/openrouter/kimi-k2.6 - 1m 33s (18%) 2. hermes/rtx5090/gemma-4-26b-a4b-nvfp4 - 2m 24s (47%) 3. opencode/rtx5090/gemma-4-26b-a4b-nvfp4 - 3m 21s (14%) 4. opencode/openrouter/deepseek-v4.1-flash - 10m 56s (96%) 5. opencode/openrouter/kimi-k2.6 - 12m 28s (53%) 6. codex - 12m 36s (73%) 7. Codex gpt-5 - 13m 02s (92%) 8. openclaw/openrouter/deepseek-v4.1-flash - 13m 23s (88%) 9. pi/openrouter/deepseek-v4.1-flash - 13m 36s (94%) 10. Claude Code - 14m 26s (100%) For a better way to browse the results, I'll attach a comment in the comment.
document fetched 2026-09-29 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1ws6rfg/harness_and_model_combinations_which_one_is_the