Harnesses
A harness is the agent program that runs a model: it reads the task, runs code and tools in the sandbox, and passes back what they return. An agent is a harness plus a model, so the same model can score differently in different harnesses. These are the harnesses we benchmark, with every model we run in each.
Every harness we benchmark
| Harness | Made by | Version | Models | Entries (ranked) | Best entry | Geospatial Agent Index | Rank |
|---|---|---|---|---|---|---|---|
| Claude Code | Anthropic | 2.1.286 | 14 | 26 (25) | Claude Opus 5.5 (max) | 85.1 | #1 |
| Codex | OpenAI | 0.162.0 | 11 | 21 (18) | GPT-6.1 Sol (xhigh) | 81.8 | #4 |
| Kimi Code | Moonshot AI | 2.1.1 | 3 | 3 (3) | Kimi K3 (Reasoning) | 75.1 | #19 |
| Terminus-2 | Harbor, open source | 2.0.0 | 27 | 38 (34) | Kimi K3 (Reasoning) | 73.1 | #25 |
| MiMo Code | Xiaomi | 0.1.16 | 1 | 1 (1) | MiMo V2.6 Flash (Reasoning) | 71.4 | #38 |
| Qwen Code | Alibaba (Qwen) | 0.26.0 | 1 | 1 (1) | Qwen 3.8 27B (Reasoning) | 69.3 | #47 |
| MiniMax Code | MiniMax | 0.7.4 | 1 | 1 (1) | MiniMax M3 (Reasoning) | 68.2 | #51 |
Best entry in each harness
Geospatial Agent Index, Text only, 0 to 100 · Higher is better · one bar per harness, its best ranked entry
Data table: Best entry in each harness
| Model | Developer | Best entry in each harness |
|---|---|---|
| Claude Opus 5.5 (max), Claude Code | Anthropic | 85 |
| GPT-6.1 Sol (xhigh), Codex | OpenAI | 82 |
| Kimi K3 (Reasoning), Kimi Code | Moonshot AI | 75 |
| Kimi K3 (Reasoning), Terminus-2 | Moonshot AI | 73 |
| MiMo V2.6 Flash (Reasoning), MiMo Code | Xiaomi | 71 |
| Qwen 3.8 27B (Reasoning), Qwen Code | Alibaba (Qwen) | 69 |
| MiniMax M3 (Reasoning), MiniMax Code | MiniMax | 68 |
Same model, different harness
| Model | Claude Code | Codex | Kimi Code | Terminus-2 |
|---|---|---|---|---|
| DeepSeek V4.1 Flash (max) | 76.1 | 77.8 | – | – |
| DeepSeek V4 Pro (high) | – | 73.3 (provisional) | – | 70.3 |
| DeepSeek V4 Flash (high) | 59.3 | 66.1 | – | 72.4 |
| Kimi K3 (Reasoning) | – | – | 75.1 | 73.1 |
| Kimi K2.7 Code (Reasoning) | – | – | 72.3 | 71.8 |
| Kimi K2.6 (Reasoning) | – | – | 70.0 | 69.6 |
| gpt-oss-120b (high) | – | 42.8 | – | 45.6 |
| gpt-oss-20b (high) | – | 9.3 | – | 33.5 |
| GLM-5.3 (max) | 74.5 | – | – | 72.4 |
| GLM-5.2 (max) | 73.5 | – | – | 71.7 |
| GLM-5.3-Flash (max) | 72.1 | – | – | 62.7 |
| GLM-4.7-Flash (Reasoning) | 41.3 | – | – | 43.2 |
Geospatial Agent Index, Text only. Bold: the best harness for that model by 3 points or more; smaller gaps are within noise. Provisional: not yet ranked. MiMo Code and Qwen Code and MiniMax Code run only their developer's model so far, so no row has a score there.