Compared with the Preview, the release version has improved significantly in agentic tasks and coding. According to benchmarks, it now outperforms GLM-5.2 and even the flagship DeepSeek V4 Pro, while retaining the same architecture and parameter count. The API has also been fully adapted for use with the Codex agent.
In the Artificial Analysis ranking, it is now just one point behind GLM-5.2 and GPT-5.6 Luna, and on par with Gemini 3.6 Flash.













