GLM-4.6

2025-09-30 · Publisher : Z.ai · Domain : Artificial Intelligence

Z.ai has released GLM-4.6, an update to GLM-4.5: context raised from 128K to 200K tokens, better agentic code, reasoning with tools, open weights under the MIT license.

### Context GLM-4.6 comes out one day after Claude Sonnet 4.5 (September 29, 2025) and two months after GLM-4.5. The open landscape includes DeepSeek-V3.1-Terminus, against which the vendor compares itself. The aim is to consolidate Z.ai's agentic code offering, sold through a GLM Coding Plan compatible with Claude Code, Cline, Roo Code and Kilo Code. ### What the model brings - A context window widened from 128K to 200K tokens, with 128K tokens of maximum output (API documentation). - Better code scores and better behavior in real agents (Claude Code, Cline, Roo Code, Kilo Code), including for polished front-end pages. - Stronger reasoning, with tool calling during inference. - Better-performing search and tool agents, and writing better aligned with human preferences. - About 15% fewer tokens than GLM-4.5 to finish a task, according to the vendor (CC-Bench). ### Architecture and specifications - Same family and same inference method as GLM-4.5 (MoE 355B total, 32B active according to the GLM-4.5 model card); the HF metadata count gives about 356.8 billion parameters. - Context: 200K; maximum output: 128K; modality: text; knowledge cutoff: not disclosed. - License: MIT. Weights: zai-org/GLM-4.6 (and FP8). No dedicated technical report: the model card points to the GLM-4.5 report (arXiv 2508.06471). ### Cost and access | Item | Value (current pricing) | |---|---| | GLM-4.6 API, input | 0.6 USD per million tokens | | GLM-4.6 API, cached input | 0.11 USD per million tokens | | GLM-4.6 API, output | 2.2 USD per million tokens | | Self-hosting | Same configuration as GLM-4.5 (8 H100 in FP8 according to its model card); not requalified for GLM-4.6 | ### Benchmarks | Benchmark | GLM-4.6 | Claude Sonnet 4.5 | GPT-5-High | |---|---|---|---| | SWE-bench Verified | 68.0 | 77.2 | 74.9 | | Terminal Bench 2.0 | 24.5 | 42.8 | 35.2 | | GPQA-Diamond | 81.0 | 83.4 | 85.7 | | AIME 2025 | 93.9 | 87.0 | 94.6 | | LiveCodeBench v6 | 82.8 | 64.0 | 87.0 | | BrowseComp | 45.1 | 24.1 | 54.9 | Source: table from the GLM-4.7 Hugging Face model card (GLM-4.6 column), scores measured by the vendor, published in December 2025 and not in the GLM-4.6 launch post, whose table is an image. On CC-Bench (52 tasks, human evaluators), GLM-4.6 obtains 48.6% wins against Claude Sonnet 4 (official post). ### Versus closed models The reference closed model of the period is Claude Sonnet 4.5, published the day before (date taken from the Anthropic announcement, not consulted again for this entry). In the vendor's table, GLM-4.6 trails Sonnet 4.5 in agentic code (68.0 versus 77.2 on SWE-bench Verified, 24.5 versus 42.8 on Terminal Bench 2.0) but is ahead of Sonnet 4.5 on BrowseComp (45.1 versus 24.1) and on AIME 2025 (93.9 versus 87.0). Against GPT-5-High, the model is below on all the lines cited. On CC-Bench, the vendor announces near-parity with Claude Sonnet 4 (48.6% wins), a model from the previous generation. The cost argument is strong: the vendor puts forward a GLM Coding Plan at one seventh of the price of an equivalent Claude offer with three times the quota (unverified commercial claim). The open choice is justified mainly by sovereignty and self-hosting, not by a score advantage. ### Key takeaways GLM-4.6 consolidates GLM-4.5 with 200K of context and fewer tokens consumed. It remains below the best closed models in agentic code, but with MIT weights and a low API price.

Tags : LLM, Open Source, Z.ai, Zhipu AI, GLM, MIT, MoE, Code