Z.ai's GLM-4.6 Matches Claude Sonnet 4 Coding at 10x Lower Cost
Z.ai's open-weight 357B MoE stretches context to 200K, closes the gap with Claude Sonnet 4 on real-world coding, and stays MIT licensed.
Takeaways
- Z.ai released GLM-4.6, an MIT-licensed 357B-parameter MoE model with 32B active parameters.
- Context window expands from 128K to 200K tokens, matching Claude Sonnet 4's window.
- Reaches 48.6% win rate against Claude Sonnet 4 on Z.ai's real-world CC-Bench coding suite.
- LiveCodeBench v6 score jumps from 63.3% (GLM-4.5) to 82.8%, matching Claude 4 tier.
- API priced at $0.43/M input and $1.75/M output tokens via Z.ai.
- Ships with native tool use, MTP speculative decoding, and integrations for Claude Code, Cline, Roo Code.
Z.ai has pushed out GLM-4.6, the newest iteration of its flagship open-weight model, and it lands squarely in the conversation about which open models can credibly replace Claude for day-to-day coding work. The weights are on Hugging Face under an MIT license, the API is live, and the pitch is straightforward: near-Sonnet quality at a fraction of the price.
What actually changed
The context window has grown from 128K to 200K tokens, matching Claude Sonnet 4 and unlocking whole-repo prompting in a single call. On top of that, Z.ai claims across-the-board gains in coding, tool use, search agents, and writing quality.
The upgrades over GLM-4.5 break down like this:
- Higher scores on code benchmarks and better real-world performance in applications such as Claude Code, Cline, Roo Code, and Kilo Code, including improvements in generating visually polished front-end pages.
- Improved reasoning performance with support for tool use during inference.
- Stronger results on tool-using and search-based agents, and more effective integration within agent frameworks.
- Better alignment with human preferences in style and readability, and more natural role-playing.
The coding jump is the real story
Z.ai ran what it calls CC-Bench, a suite of practical coding tasks executed inside the Claude Code environment. Across 74 real-world tests, GLM-4.6 surpasses Claude Sonnet 4 and other domestic models, and its average token consumption runs over 30% below GLM-4.5, the lowest among comparable models. On CC-Bench it reaches near parity with Claude Sonnet 4 at a 48.6% win rate.
Public benchmarks show an even sharper delta over the previous generation. On LiveCodeBench v6, which involves writing, executing, and debugging code across languages, GLM-4.6 scored 82.8%, a large jump from GLM-4.5's 63.3% and roughly matching Anthropic's Claude 4. Across eight benchmarks including AIME 25, GPQA, LCB v6, HLE, and SWE-Bench Verified, GLM-4.6 performs on par with Claude Sonnet 4 and 4.6 on several leaderboards.