Model Drop

GLM-5

February 11, 2026 2 min read

GLM-5, which Z.ai released on February 11, 2026, is a 744B-parameter open-weights model built for long agentic coding and systems engineering work.

Model: GLM-5 (zai-org/GLM-5)
Maker: Z.ai
Ship date: February 11, 2026
Pricing: $1 / $3.20 per M tokens (Z.ai API); free weights
Context: 200K tokens
Weights: Open (MIT)
Where you can use it:
- Free chat at Z.ai and the Z.ai API
- MIT weights on Hugging Face
When to reach for it:
- Long agentic coding and engineering tasks
- A self-hosted, MIT-licensed frontier model
- Skip it for images; it takes text only
How it compares:
- Z.ai pitches it against Claude Opus 4.5
- Open rivals: Kimi K2.5 and DeepSeek V3.2
What makes it different:
- 744B MoE with 40B active parameters
- Uses DeepSeek Sparse Attention
- Runs on Huawei Ascend and other Chinese chips
Early reactions: Hacker News users liked its coding and tool use for the price, but some doubted the benchmarks would hold up in real work and noted it costs more than GLM-4.7.
Read more: GLM-5: From Vibe Coding to Agentic Engineering, model card

The benchmarks

Z.ai's launch table runs every model in thinking mode (GLM-5 Thinking, Claude Opus 4.5 with extended thinking, Gemini 3 Pro at high thinking, GPT-5.2 at xhigh), with Kimi K2.5 and DeepSeek V3.2 as the open rivals. Humanity's Last Exam is the text-only subset, Terminal-Bench 2.0 uses the Terminus 2 harness on the original task set, and MCP Atlas is the 500-task public set.

Z.ai's numbers

SWE-bench Verified (real GitHub fixes): 77.8%. +1.0 over Kimi K2.5 (76.8%), +4.7 over DeepSeek V3.2 (73.1%), −3.1 under Claude Opus 4.5 (80.9%), −2.2 under GPT-5.2 (80.0%).
Terminal-Bench 2.0 (agentic terminal tasks): 56.2%. −3.1 under Claude Opus 4.5 (59.3%), +2.0 over Gemini 3 Pro (54.2%), +2.2 over GPT-5.2 (54.0%), +5.4 over Kimi K2.5 (50.8%).
BrowseComp (hard web research): 62.0%. +1.4 over Kimi K2.5 (60.6%), +10.6 over DeepSeek V3.2 (51.4%), +25.0 over Claude Opus 4.5 (37.0%), +24.2 over Gemini 3 Pro (37.8%).
MCP Atlas (multi-tool MCP tasks): 67.8%. −0.2 under GPT-5.2 (68.0%), +1.2 over Gemini 3 Pro (66.6%), +2.6 over Claude Opus 4.5 (65.2%), +4.0 over Kimi K2.5 (63.8%).
Humanity's Last Exam (with tools) (expert questions): 50.4%. −1.4 under Kimi K2.5 (51.8%), +9.6 over DeepSeek V3.2 (40.8%).
Humanity's Last Exam (no tools) (expert questions): 30.5%. −1.0 under Kimi K2.5 (31.5%), +2.1 over Claude Opus 4.5 (28.4%), −6.7 under Gemini 3 Pro (37.2%), −4.9 under GPT-5.2 (35.4%).
GPQA Diamond (PhD-level science): 86.0%. −1.6 under Kimi K2.5 (87.6%), −1.0 under Claude Opus 4.5 (87.0%), −5.9 under Gemini 3 Pro (91.9%), −6.4 under GPT-5.2 (92.4%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2, AA estimate): 27.9. +4.4 over Kimi K2.5 (23.5), +6.4 over DeepSeek V3.2 (21.5), −1.2 under Claude Opus 4.5 (29.1).