Benchmarking GLM-5.2 vs Opus 4.8 for real-world long-context retrieval - Evals - Braintrust

Our benchmark shows that GLM-5.2 is slightly behind Opus 4.8 on retrieval, while average provider cost per trace is about 76-78% lower across the two context tiers. On a perturbation control slice, GLM is about 86% cheaper per correct answer.