# CyberGym → MCP Atlas

Calibration · https://benchgap.net/calibration/cybergym/current/to/mcpatlas/current · data retrieved 2026-11-18

MCP Atlas is estimated from CyberGym with a Hill curve fitted on 12 models measured on both: y = 0.0000 + (0.8126 − 0.0000)·x^6.00 / (0.46036^6.00 + x^6.00), R² = 0.75, cross-validated error 11.1 pp. It is used for 4 estimates.

| Estimated model | CyberGym | MCP Atlas | Source |
|---|---|---|---|
| [Fugu Cyber](https://benchgap.net/model/sakana-fugu-cyber.md) | 86.9% | 79.5% | estimated ± 11.1 pp, low confidence |
| [Gemini 3.8 Flash Cyber](https://benchgap.net/model/gemini-3-8-flash-cyber.md) | 86.2% | 79.4% | estimated ± 11.1 pp, low confidence |
| [Gemini 3.5 Flash Cyber](https://benchgap.net/model/gemini-3-5-flash-cyber.md) | 83.2% | 79.0% | estimated ± 11.1 pp, low confidence |
| [Claude Mythos Preview](https://benchgap.net/model/claude-mythos-preview.md) | 83.1% | 79.0% | estimated ± 11.1 pp, low confidence |

## Fitted on

| Model | CyberGym | MCP Atlas |
|---|---|---|
| [Claude Opus 4.5](https://benchgap.net/model/claude-opus-4-5.md) | 50.6% | 42.3% |
| [Claude Opus 4.7](https://benchgap.net/model/claude-opus-4-7.md) | 73.1% | 77.3% |
| [DeepSeek V4 Flash 0731](https://benchgap.net/model/deepseek-v4-flash-0731.md) | 76.7% | 69.0% |
| [DeepSeek V4 Pro](https://benchgap.net/model/deepseek-v4-pro.md) | 83.3% | 73.6% |
| [GLM-5](https://benchgap.net/model/glm-5.md) | 43.2% | 31.1% |
| [GLM-5.1](https://benchgap.net/model/glm-5-1.md) | 68.7% | 71.8% |
| [GPT-5.4](https://benchgap.net/model/gpt-5-4.md) | 79.0% | 70.6% |
| [GPT-5.5](https://benchgap.net/model/gpt-5-5.md) | 81.8% | 75.3% |
| [Hy4 preview](https://benchgap.net/model/hy4-preview.md) | 78.4% | 83.7% |
| [Ling 3.1 Flash](https://benchgap.net/model/ling-3-1-flash.md) | 87.9% | 83.0% |
| [Muse Spark 1.1](https://benchgap.net/model/muse-spark-1-1.md) | 59.0% | 88.1% |
| [Step 5 Preview](https://benchgap.net/model/step-5-preview.md) | 84.7% | 85.6% |
