CL-bench
Learn new rules, procedures and domain knowledge from context, then apply them.
Agent Index weight
5%
in the Security & context learning pillar
Models tested
19
from the benchmark’s leaderboard and reports
Results
| # | Model | Score | Normalised | Harness | Cost/task | Date |
|---|---|---|---|---|---|---|
| 1 | GPT-5.4 Supported OpenAI | 27.9% | source | — | — | |
| 2 | GPT-5.1 Supported OpenAI | 23.7% | source | — | — | |
| 3 | Grok 4.20 (Non-Reasoning) Estimated xAI | 22.2% | source | — | — | |
| 4 | Claude Opus 4.5 (latest) Supported Anthropic | 21.1% | source | — | — | |
| 5 | Gemini 3.1 Pro Preview Supported Google | 20.8% | source | — | — | |
| 6 | Claude Opus 4.6 Supported Anthropic | 20.7% | source | — | — | |
| 7 | Qwen3.6 Plus Estimated Alibaba (Qwen) | 20.3% | source | — | — | |
| 8 | Qwen3.5 Plus Estimated Alibaba (Qwen) | 19.8% | source | — | — | |
| 9 | Kimi K2.5 Supported Moonshot AI | 19.3% | source | — | — | |
| 10 | GLM-5 Supported Z.ai (Zhipu) | 18.7% | source | — | — | |
| 11 | GPT-5.2 Supported OpenAI | 18.2% | source | — | — | |
| 12 | o3 Supported OpenAI | 17.8% | source | — | — | |
| 13 | Kimi K2 Thinking Estimated Moonshot AI | 17.6% | source | — | — | |
| 14 | GLM-4.7 Estimated Z.ai (Zhipu) | 15.9% | source | — | — | |
| 15 | Gemini 3 Pro Preview Supported Google | 15.8% | source | — | — | |
| 16 | MiMo V2 Pro Free Xiaomi | 15.7% | source | — | — | |
| 17 | Qwen3 Max Estimated Alibaba (Qwen) | 14.5% | source | — | — | |
| 18 | DeepSeek V3.2 Supported DeepSeek | 13.2% | source | — | — | |
| 19 | MiniMax-M2.5 Estimated MiniMax | 11.4% | source | — | — |
For models without a CL-bench result, the index estimates one from Epoch’s Capabilities Index using a logistic fit on 17 models (residual SD 0.17 on the logit scale). Estimates are labelled and widen the model’s range.