GLM-4.6

GLM-4.6 has credible official coverage from Z.ai documentation and the main product page, which position it as an open-weight model for coding, long-context reasoning, search, writing, and agentic development. The strongest public evidence is in coding: supplied sources report SWE-bench Verified at 68.0%, LiveCodeBench at 70.1, Aider-polyglot at 39.1, and Terminal-Bench 2.0 accuracy of 24.5%±2.4 when used in a Terminus 2 setup. DeepSWE was checked, but GLM-4.6 was not listed, so no DeepSWE result should be inferred. Artificial Analysis provides useful third-party performance and pricing context: Intelligence Index 30, about 54.9 output tokens/s, and provider benchmarking showing Novita at 61.5 tokens/s with 1.39s TTFT. Listed pricing of $0.60/M input and $2.20/M output tokens is competitive for a coding-capable open-weight model, though official pricing was not found in the supplied Z.ai sources. Multimodal evidence is limited; the supplied positioning is mainly text, coding, search, and agent workflows. Public preference signal is present but not frontier-level. Arena lists glm-4.6 around rank 64 overall, with lower ranks for coding and hard prompts, while BenchLM reports Arena Elo and preference slices but notes limited sourced benchmark coverage. Overall, GLM-4.6 looks like a cost-effective, accessible coding and agentic-work model with decent public benchmark support, but evaluation confidence is constrained by mixed source types, missing official pricing/license details, and absent DeepSWE coverage.