Key findings
- Changing only the harness moved a model's quality score by 25 points on average and by up to 61 points (GLM-5.3).
- Cost moved more than quality: the same model cost up to 3.6 times as much on one harness as on another.
- No harness was best for every model. Claude Code was best for 7 of the 13 models it ran, OpenCode for 6 of 18 and Codex for 5 of 14.
- A vendor's own harness was not always the best home for its model. Kimi K3 scored lowest on Kimi CLI and DeepSeek Flash lowest on the DeepSeek harness.
What was compared
The five harnesses on the board are Claude Code, OpenCode, Codex, Kimi CLI and the DeepSeek harness. OpenCode, Claude Code and Codex ran models from several vendors; Kimi CLI and the DeepSeek harness ran their own vendors' models.
All figures below come from the overall board, where every pairing is scored on the same Elo-like scale. A difference of 40 points means the higher pairing is expected to win about 56% of comparisons on the same task.
Quality moves by tens of points
For each model, the table shows its best and worst harness and the gap between them.
| Model | Best harness | Worst harness | Quality gap |
|---|---|---|---|
| GLM-5.3 | OpenCode 993 | Codex 932 | 61 |
| MiniMax M2.5 | OpenCode 700 | Codex 641 | 59 |
| Kimi K3 | OpenCode 1147 | Kimi CLI 1094 | 53 |
| GLM-5.3 Prime | Claude Code 1021 | OpenCode 980 | 41 |
| DeepSeek Flash | Claude Code 1039 | DeepSeek harness 1000 | 39 |
| Claude Opus 5.5 | Claude Code 1335 | OpenCode 1298 | 37 |
| GPT-6 Sol | OpenCode 1190 | Codex 1155 | 35 |
| Qwen3.8 Flash | Codex 859 | Claude Code 831 | 28 |
| GPT-6 Astra | Codex 1235 | OpenCode 1208 | 27 |
| DeepSeek V4 Pro | Claude Code 914 | Codex 892 | 22 |
| Qwen3.7 Plus | Codex 856 | Claude Code 835 | 21 |
| GPT-6.1 Sol | OpenCode 1196 | Codex 1187 | 9 |
| GPT-5.6 Terra | OpenCode 1105 | Codex 1098 | 7 |
| Claude Sonnet 5 | Claude Code 906 | OpenCode 900 | 6 |
| Qwen3.8 Max | Codex 1001 | Claude Code 998 | 3 |
| Claude Haiku 4.5 | Claude Code 681 | OpenCode 678 | 3 |
| Claude Fable 5.1 | Claude Code 1320 | OpenCode 1318 | 2 |
| GPT-5.6 Sol | Codex 1174 | OpenCode 1173 | 1 |
The average gap is 25 points and the median is 27.5. For seven of the 18 models the harness made almost no difference, under 10 points. For another seven it made 35 points or more.
A gap of this size can reorder models. GPT-6 Sol on Codex scores 1155, below the older GPT-5.6 Sol on the same harness at 1174. Moved to OpenCode, GPT-6 Sol scores 1190 and is ahead again.
Cost and time move even more
The harness decides how much context goes to the model, how often it calls tools and when it stops. That shows up in spend and duration more than in quality:
- MiniMax M2.5 costs $0.55 a task on Claude Code and $0.15 on Codex, 3.6 times as much, for a quality difference of 38 points in Claude Code's favour.
- Claude Sonnet 5 scores 906 on Claude Code and 900 on OpenCode, but costs $1.62 against $0.91.
- Kimi K3 costs $2.46 a task on Claude Code and $1.23 on OpenCode, and scores 52 points lower on Claude Code.
- Qwen3.8 Max scores between 998 and 1001 on all three harnesses it ran, but takes 1800.6 seconds a run on Codex against 1253.1 on OpenCode, and costs twice as much there.
In several cases the harness that scored higher was also the cheaper and faster one. Claude Opus 5.5 on Claude Code scores 37 points more than on OpenCode while costing $2.60 instead of $3.17 and finishing in 746.5 seconds instead of 1065.0.
No harness wins everywhere
| Harness | Models run | Best harness for | Average gap to the best |
|---|---|---|---|
| Claude Code | 13 | 7 | 10.6 |
| OpenCode | 18 | 6 | 10.4 |
| Codex | 14 | 5 | 19.2 |
| DeepSeek harness | 2 | 0 | 30.0 |
| Kimi CLI | 1 | 0 | 53.0 |
Claude Code and OpenCode are close on average, each within about 10 points of the best harness for the models they ran. Codex trails by 19 points on average, but it is the best harness for GPT-6 Astra, GPT-5.6 Sol and three Qwen models.
Models do not always do best on their vendor's harness. Claude Opus 5.5 and GPT-6 Astra do, by 37 and 27 points. Kimi K3 scores 1094 on Kimi CLI and 1147 on OpenCode. DeepSeek Flash scores 1000 on the DeepSeek harness and 1027 to 1039 on the other three. The vendor harnesses do keep an edge on speed and price in some cases: DeepSeek V4 Pro is cheapest and fastest on the DeepSeek harness, at $0.18 and 488.5 seconds a run.
What this means when choosing an agent
- Pick the pairing, not just the model. The model page on Work Arena lists every harness a model ran on and where each one placed.
- Check cost per harness before committing. Two harnesses with the same quality can differ by a factor of two in spend.
- Do not assume the vendor's harness is the best one. For several models a general-purpose harness did better.
Every model and harness has its own page under models and harnesses, and the leaderboard shows the pairing matrix for the overall board and for each field.
Questions
Does the agent harness matter as much as the model
Usually less, but it is not small. Across 18 models, switching harness moved quality by 25 points on average. The gap between neighbouring models in the top twenty is often smaller than that, so a better harness can be worth a place or two.
Is Claude Code better than Codex
It depends on the model. Nine models ran on both. Claude Code scored higher for DeepSeek Flash, DeepSeek V4 Pro, GLM-5.3, GLM-5.3 Prime and MiniMax M2.5. Codex scored higher for Kimi K3 and the three Qwen models. Claude models were not run on Codex, and GPT models were not run on Claude Code.
Figures in this article come from the leaderboard data of 2 November 2026. The leaderboard shows any changes since.