Stage#1
Unified Base Model
GLM-5 uses a single unified model base, eliminating the need for separate models when switching between think and non-think usage.
GLM-5 MODEL HUB
GLM-5: Benchmark, Price, API and Architecture
Evaluate GLM-5 capability ceilings, compare benchmark scores, review API pricing, and explore the architecture — all in one place.
TOTAL PARAMETERS
744B
ACTIVE MoE
40B
CONTEXT WINDOW
200K
LICENSE
MIT
Community
What People Are Saying About GLM-5
Watch these in-depth reviews and tutorials from the AI community showcasing GLM-5's capabilities.
GLM-5 Fully Tested: I GOT EARLY ACCESS & YES, IT BEATS 4.6 OPUS!
Full test covering coding quality, multi-step task behavior, and direct model comparisons on agentic coding benchmarks.
AICodeKing
44.2K views · Feb 11, 2026
GLM-5 Is HERE – Is THIS the BEST Open Source Coding Model?
Hands-on demo with browser automation, game generation, and Python 3D tasks to evaluate real-world coding output quality.
Bijan Bowen
15K views · Feb 11, 2026
GLM-5: From Vibe Coding to Agentic Engineering
Use-case walkthrough from quick coding prompts to agentic engineering pipelines, exploring GLM-5 real-world deployment paths.
Fahd Mirza
3.5K views · Feb 11, 2026
Overview
What Is GLM-5
GLM-5 is Zhipu AI's fifth-generation large language model, targeting complex systems engineering and long-horizon agentic tasks. It scales up from GLM-4.5 with 355B/32B active parameters to 744B/40B active, and grows pretraining data from 23T to 28.5T tokens.
Unified Hybrid Reasoning
Think and non-think modes share one base model, with differences mainly in post-training strategy.
Long Context with Cost Controls
DeepSeek Sparse Attention reduces inference cost while preserving 200K context capability for long-document workflows.
Open Source + API Access
GLM-5 weights are available on Hugging Face and ModelScope under MIT license, with API access through multiple providers.
Architecture
GLM-5 Architecture
GLM-5 uses a unified model architecture with hybrid reasoning mode switching. Think and non-think share the same base, with DSA for long-context efficiency and the slime asynchronous RL infrastructure for training.
1
744B
Total Parameters
2
40B
Active Parameters
3
28.5T Tokens
Pretraining Data
4
Unified Think/Non-think
Reasoning Mode
Stage#2
Post-Training Specialization
Differences between think and non-think behavior are introduced in post-training, enabling quality and latency tradeoffs.
Stage#3
Hybrid Reasoning Switch
Hybrid reasoning can be enabled based on task complexity and workflow needs for flexible deployment.
Stage#4
DSA for Long Context
DeepSeek Sparse Attention reduces inference cost while preserving long-context capability across large input sequences.
Stage#5
Asynchronous RL Infrastructure
The slime framework provides asynchronous reinforcement learning for higher training throughput and faster iteration cycles.
Stage#6
Function Calling + Parallel Tools
GLM-5 supports Function Calling and parallel tool calls, enabling complex agentic workflows with multi-step orchestration.
Context Window
GLM-5 Context Window
Context window and output limits for GLM-5 across major providers. Displayed limits can vary by platform and routing layer.
1
200K context / up to 128K output
GLM-5 (docs.z.ai)
2
202,752 context
GLM-5 (OpenRouter)
3
up to 202,752
OpenRouter max completion
4
Limits vary by endpoint
Provider caveat
Stage#1
Long Docs and Multi-File Tasks
A 200K context window supports long-document QA, cross-file code analysis, and multi-step planning workflows.
Stage#2
Controllable Output Limits
GLM-5 supports up to 128K output tokens in official docs. Effective output limits can vary by provider and endpoint settings.
Stage#3
Platform-Level Differences
The 202,752 context number comes from OpenRouter. For production, verify effective limits on your target provider.
Benchmark
GLM-5 Benchmark Overview
Published benchmark visuals covering agentic, reasoning, coding, and long-horizon tasks for side-by-side model comparison.
Sources: docs.z.ai and z.ai/blog/glm-5, captured February 12, 2026
GLM-4.7GLM-5Claude Opus 4.5Gemini 3.0 ProGPT-5.2
LLM Performance Evaluation
Eight public benchmarks including Humanity's Last Exam, SWE-bench, Terminal-Bench, MCP-Atlas, and Vending Bench 2.
CC-Bench-V2
Comparison chart for frontend, backend, and long-horizon engineering tasks.
77.8
SWE-bench Verified
GLM-5 score on SWE-bench Verified for real-world code repair tasks.
73.3
SWE-bench Multilingual
Performance on multilingual software engineering tasks across multiple languages.
56.2
Terminal-Bench 2.0
Score on terminal-agent workloads measuring command-line task completion.
$4,432
Vending Bench 2
Final balance achieved by GLM-5 in the business simulation benchmark.
API Price
GLM-5 API Price by Platform
Pricing from docs.z.ai and OpenRouter as of February 12, 2026. All prices shown here are in USD; re-check provider pages before production launch.
1
$1.00 / 1M tokens
GLM-5 Input
2
$0.20 / 1M tokens
GLM-5 Cached Input
3
$3.20 / 1M tokens
GLM-5 Output
4
$1.20 in / $5.00 out
GLM-5-Code
Stage#1
GLM-5 via docs.z.ai
Input $1.00 / 1M, cached input $0.20 / 1M, and output $3.20 / 1M in USD.
Stage#2
GLM-5-Code via docs.z.ai
Input $1.20 / 1M, cached input $0.30 / 1M, and output $5.00 / 1M in USD.
Stage#3
Cached Input Storage
Cached input storage is marked as limited-time free on the official pricing page.
Stage#4
GLM-5 via OpenRouter
GLM-5 is listed at $1 / 1M input and $3.20 / 1M output in USD.
Stage#5
Per-1M Token Units
Both docs.z.ai and OpenRouter list prices per 1M tokens; check provider billing granularity before launch.
Stage#6
Provider Billing Differences
Final billed cost may differ by routing policy, platform markup, and cache behavior across providers.
Benefits
Why Teams Are Evaluating GLM-5 in 2026
GLM-5 stands out for systems-engineering workloads, long-horizon agent tasks, competitive pricing, and record-low hallucination rates.
GLM-5 is positioned for complex systems engineering and high-complexity execution workflows with multi-step tool use.
FAQ
GLM-5 Frequently Asked Questions
Common questions about GLM-5 benchmarks, API pricing, context windows, and model capabilities.
What is the GLM-5 API price?
As of February 12, 2026, docs.z.ai lists GLM-5 at $1.00/1M input, $0.20/1M cached input, and $3.20/1M output; GLM-5-Code is $1.20/1M input, $0.30/1M cached input, and $5.00/1M output. OpenRouter lists GLM-5 at $1/1M input and $3.20/1M output.
What is the GLM-5 context window?
GLM-5 supports a 200K context window with up to 128K output tokens in official docs. OpenRouter currently lists 202,752 context, and effective limits can vary by provider endpoint.
Is GLM-5 a Mixture-of-Experts model?
Yes. GLM-5 has 744B total parameters with 40B active per token, using a sparse-activation MoE architecture with 256 experts and 8 activated per token.
How does GLM-5 compare to other frontier models?
GLM-5 scores 77.8 on SWE-bench Verified. In published comparison charts, Claude Opus 4.5 is 80.9 and Gemini 3.0 Pro is 76.2. GLM-5 ranks first among open-weight models on Vending Bench 2.
Why do OpenRouter prices differ from api.z.ai?
OpenRouter is a routing platform and may apply platform-level routing and margin policies. Always verify final billing rules and units on each provider page.
Are GLM-5 weights open-source?
Yes. GLM-5 weights are available on Hugging Face and ModelScope under the MIT License, supporting local deployment with vLLM and SGLang.