HomeAI models
Z.ai: GLM 5
Z.AI
•
GLM-5
•
Released 
February 2026

Z.AI
GLM-5

Open-weight 744B MoE reasoning model built for long-horizon agentic coding and complex systems engineering.

Modality:
Text
PDF
model ID
z-ai/glm-5
Start for free
arrow black icon
View API Docs
arrow white icon

Output Speed *

N/A
tok/s

Intelligence Index *

27.9
/ 100

Context Window *

202752
tokens

Input price

3.6
Anytoken

Output price

11.52
Anytoken
GLM-5: An Open-Weight Reasoning Model for Long-Horizon Agentic Engineering GLM-5 is Z.ai's flagship open-weight reasoning model, built on a 744B-parameter Mixture-of-Experts architecture activating roughly 40B parameters per token. It targets complex systems engineering and multi-step agent workflows rather than simple single-turn generation. Reasoning runs always-on with selectable effort levels, and the model pairs strong coding and agentic benchmark results with an MIT license. Teams building autonomous coding agents, repository-scale refactoring tools, and tool-using systems that need a self-hostable frontier-class open model benefit most from GLM-5. Start building with the GLM-5 API on AnyAPI.ai.

Performance

Where GLM-5 Earns Its Place: Long-Horizon Agentic Coding

GLM-5 is tuned for end-to-end software engineering and multi-step agent execution rather than quick completions. On Z.ai's reported benchmarks it reaches 75.9 on SWE-bench Verified and became the first open-weight model to score 50 on the Artificial Analysis Intelligence Index v4.0, an 8-point jump over GLM-4.7 driven largely by agentic and knowledge gains. For production, this means GLM-5 can sustain long tool-using trajectories and repository-scale changes while remaining self-hostable under an MIT license, making it a credible open alternative to closed frontier models for coding agents.

Benchmarks

GLM-5 Benchmarks: Reasoning, Coding, and Agentic Evidence

Independent evaluation from Artificial Analysis placed GLM-5 at 50 on the Intelligence Index v4.0, the first open-weight model to reach that threshold. Z.ai-reported figures show strong reasoning (AIME 2026 I 92.7, GPQA-Diamond 86.0) and competitive agentic coding (SWE-bench Verified 75.9, with HLE 30.5 rising to 50.4 with tools). On LMArena's Text Arena it ranked first among open models with a score around 1452. Throughput is independently measured at roughly 106 tokens/second — adequate for agent loops but slower than lighter inference-optimized models.

Output Speed

*
N/A
tok/s

Intelligence Index

*
27.9
/ 100

MMLU *

Broad world knowledge and problem-solving
0
%

GPQA *

PhD-level scientific reasoning across physics, biology, chemistry.
82
%

HLE *

Adherence to multi-step structured instructions.
29
%

LiveCodeBench *

Tool-calling reliability in long agentic loops.
0
%

Technical Specifications

What the model supports

GLM-5 is a text-in, text-out reasoning model served with a 204,800-token context window and up to 128,000 output tokens. The most consequential detail is that reasoning is always enabled and cannot be switched off — the thinking trace and the final answer share the same output budget, so max_tokens must be set generously and effort tuned deliberately. It exposes OpenAI-compatible function calling and JSON-schema structured outputs, and ships as open weights under the MIT license, enabling self-hosting and fine-tuning.
Verified Specifications — 
GLM-5
*
Input modalities
Text
PDF
output modalities
Text
Context window
202752
 tokens
Maximum output tokens
131072
Reasoning
Yes
Knowledge cutoff
February 2026
Pricing (standard)
3.6
 AnyTokens in
 / 
11.52
 AnyTokens out

Comparison

GLM-5 vs GLM-4.7: What You Gain Moving Up the Family

GLM-5 is the direct successor to GLM-4.7 in Z.ai's open-weight lineage, and both target coding and agentic workloads under permissive licensing. GLM-5 scales the architecture substantially — from 355B to 744B total parameters — and expands pre-training data to 28.5T tokens. The practical decision is whether your workload needs GLM-5's higher reasoning and agentic ceiling, or whether GLM-4.7's smaller footprint is sufficient and cheaper to self-host for your task mix.

Dimension
GLM-5
GLM-4.7 (Reasoning)
Context window *
202752
tokens
202752
tokens
Output speed *
N/A
tok/s
0.00
tok/s
Intelligence Index *
27.9
22.2
Input pricing
3.6
AnyToken
3.6
AnyToken
Output pricing
11.52
AnyToken
13.2
AnyToken
Knowledge cutoff *
February 2026
December 2025

Choose GLM-5 when you need the strongest open-weight reasoning and long-horizon agentic coding available in the family — it posts an 8-point Intelligence Index gain over GLM-4.7 and higher SWE-bench Verified results. Choose GLM-4.7 when inference cost, memory footprint, or simpler single-turn tasks dominate, since its smaller size is easier and cheaper to run and reasoning overhead is lower. For high-volume routine generation, GLM-5's always-on reasoning may be unnecessary.

Limitations & Trade-offs

Where GLM-5 falls short

1
Always-on reasoning overhead. GLM-5 cannot disable reasoning; the thinking trace consumes part of the shared 128K output budget on every call. For trivial transformations, classification, or high-volume short responses, this inflates output token usage and latency even at low effort. If your workload is mostly simple single-turn generation, a model that allows reasoning to be fully turned off is more economical and faster.
2
Context smaller than later siblings. The GLM-5 endpoint provides a 204,800-token window, materially smaller than the 1M-token context offered by later GLM-5.x releases. For whole-repository ingestion or very long document sets in a single pass, this cap may force chunking or retrieval. Workloads that depend on million-token context should evaluate a newer GLM-5.x variant instead.
3
Text-only modality. GLM-5 accepts and produces text only — it has no image, audio, or video input. Multimodal agent tasks such as screenshot-driven UI work or visual debugging require a different model. Do not assume multimodal capability from sibling models in the family; those capabilities arrived in separate later releases, not the original GLM-5.
4
Moderate throughput and heavy self-hosting requirements. Independently measured output speed is around 106 tokens/second, slower than lighter inference-optimized models, and the 744B MoE weights impose very high hardware requirements for self-hosting. Latency-sensitive real-time interfaces or teams without substantial GPU capacity may prefer a smaller, faster model or managed API access.

Best-Fit Workloads

Where this model earns its place

01

Autonomous coding agents


GLM-5 is explicitly engineered for long-horizon agentic engineering, with strong SWE-bench Verified and agentic benchmark results. It maintains coherence across extended multi-step tool-using trajectories, making it suitable as the backend for coding agents that plan, edit, and verify across many turns. Use max reasoning effort for the hardest tasks and set output budgets generously to accommodate the reasoning trace.

02

Repository-scale refactoring and bug fixing


With a 204K-token context and frontier-class open coding performance, GLM-5 handles multi-file changes that require tracking dependencies across a codebase rather than single-function completion. This fits build tooling and developer platforms that run real bug fixes and structural changes. For whole-repo ingestion exceeding the context window, pair it with retrieval or consider a larger-context GLM-5.x release.

03

Complex multi-step reasoning


Always-on reasoning and strong math and knowledge scores (AIME 2026 I 92.7, GPQA-Diamond 86.0) make GLM-5 appropriate for difficult analytical tasks, architectural planning, and problems that benefit from explicit chain-of-thought. The selectable effort control lets teams reserve deep reasoning for genuinely hard turns while using lighter effort elsewhere to manage token cost.

04

Self-hosted open-weight deployments


Released under the MIT license with open weights on Hugging Face, GLM-5 suits teams requiring on-premises or data-sovereign deployment of a frontier-class reasoning model without closed-API dependency. This is valuable for regulated environments and fine-tuning. Note the substantial GPU footprint of the 744B MoE architecture; smaller teams may prefer managed API access over self-hosting.

Pricing in anytokens via AnyAPI
Input
3.6
₳
Output
11.52
₳
Cache write
—
₳
Cache read
—
₳
Get API access
arrow black icon
Read Docs

on this page

PerformanceBenchmarksTechnical SpecificationsComparisonLimitations & Trade-offsBest-Fit WorkloadsIntegrationFAQ

Integration

Access GLM-5 via AnyAPI.ai

Access GLM-5 through AnyAPI.ai using a unified API built for multi-model AI applications. Integrate GLM-5 without maintaining a separate provider-specific connection, and keep the flexibility to test, switch, or combine models as your application requirements evolve.

01

One API integration

Access GLM-5 and other AI models through the same API workflow instead of maintaining separate integrations for every provider.

02

Easy model switching

Test GLM-5 against alternative models or switch models as your performance, capability, or cost requirements change without rebuilding your application around another provider API.

03

Flexible for production

Use GLM-5 from experimentation through production while keeping your AI stack flexible as workloads, traffic, and model requirements evolve.

04

Multi-model applications

Use GLM-5 for the workloads where it performs best and combine it with other models for tasks that require different capabilities, performance, or efficiency.

Start free
arrow black icon
Read Docs

Frequently Asked Questions

Answers to common questions about integrating and using this AI model via AnyAPI.ai

GLM-5 is Z.ai's flagship open-weight reasoning model, released on February 11, 2026. It uses a 744B-parameter Mixture-of-Experts architecture with roughly 40B active parameters per token and is designed for complex systems engineering and long-horizon agentic coding tasks. It is distributed under the MIT license.

The GLM-5 API endpoint provides a 204,800-token context window and supports up to 128,000 output tokens. Because reasoning is always enabled, the reasoning trace and the final answer share that output budget, so you should set a generous max_tokens for complex tasks.

No. GLM-5 always operates with reasoning enabled and cannot have thinking disabled. Developers instead select a reasoning effort level — low, high, or max, with max as the default — to balance reasoning depth against output token cost and latency.

Yes. GLM-5 supports OpenAI-compatible function calling through tools and tool_choice, and structured outputs via a JSON schema in response_format. It also supports streaming, which makes it suitable as a backend for tool-using coding agents.

GLM-5 is positioned as one of the strongest open-weight models for coding and agentic tasks. Z.ai reports 75.9 on SWE-bench Verified, and it was the first open-weight model to score 50 on the Artificial Analysis Intelligence Index v4.0, making it a credible open alternative to closed frontier models for agentic engineering.

* Benchmark data source: Artificial Analysis artificialanalysis.ai