GLM 5.2 is a 753B parameter MoE model from Z.ai with a 1M-token context window, MIT open-source license, and API pricing that comes in at roughly 6x less than GPT-5.5. It leads on SWE-bench Pro and MCP-Atlas, trails on Tool-Decathlon and FrontierSWE, and introduces IndexShare, an architectural change that cuts per-token FLOPs by 2.9x at long context. This post covers the full benchmark breakdown, what changed from 5.1, and whether the cost advantage holds up under scrutiny.

Z.ai launched GLM 5.2 on June 13, 2026, and published zero benchmark numbers on day one. No SWE-bench, no HumanEval, nothing. The open weights dropped a few days later, and that's when the official benchmark table finally appeared.

The numbers were worth waiting for.

This isn't a one-off quirk. Z.ai did the same thing with GLM 5.1: ship first, publish numbers later. It's worth flagging because any figures circulating before June 16 came from leaks or third-party speculation, not the official release. Everything in this post pulls from Z.ai's official announcement and repository.

GLM 5.2's 1M-token context and low per-token cost make it a strong fit for agentic workflows. Agent Builder lets you put models like this to work. Powered by Generative UI, it connects directly to your data sources and lets you build agents that respond with interactive charts, tables, and cards.

Build Your AI Agent with GLM 5.2

What Is GLM 5.2?

GLM 5.2 is a coding-focused large language model built by Z.ai, the international brand of Zhipu AI, which was spun out of Tsinghua University. It sits in the same architectural family as GLM-5 (which launched at 744B parameters), with a few meaningful upgrades under the hood.

Here's what you're working with:

Architecture: 753B parameter Mixture-of-Experts (MoE), ~40B active parameters per token
Context window: 1,000,000 tokens via glm-5.2[1m]; up to 131,072 output tokens
Training data: 28.5 trillion tokens
Thinking modes: High and Max (some providers map xhigh to Max)
License: MIT; weights on HuggingFace at zai-org/GLM-5.2 and zai-org/GLM-5.2-FP8
Modality: Text only at launch; no confirmed multimodal support
Primary focus: Coding and agentic tasks

The MIT license matters more than it might seem. It places no regional restrictions on use or distribution, a detail that's become practically significant given recent export controls (more on that below).

GLM 5.2's Biggest Architectural Bet: IndexShare

If there's one technical decision that defines GLM 5.2, it's IndexShare, and it's directly tied to why a 1M-token context window doesn't have to be ruinously expensive.

Here's the problem it solves. Standard sparse attention in MoE models computes a separate attention index for each layer. At 1M-token context lengths, that gets expensive fast. IndexShare reuses the same attention indexer across every four sparse attention layers instead of recomputing it per layer.

The result: a 2.9x reduction in per-token FLOPs at 1M context length.

GLM 5.2 also ships an improved Multi-Token Prediction (MTP) layer for speculative decoding, which raises accepted token length by up to 20%, translating to faster effective throughput at inference.

Why does this matter for developers? Think about an agentic coding workflow where the agent holds an entire large codebase in context, reads files, makes edits, runs tools, and iterates across a long session. That's exactly the scenario where per-token cost at long context has historically been the dealbreaker. IndexShare is a direct response to that constraint.

GLM 5.2 Benchmark Results*

Benchmark	GLM 5.2	GPT-5.5	Claude Opus 4.8	Verdict
SWE-bench Pro	62.1	58.6	—	GLM 5.2 leads
Terminal-Bench 2.1	81.0–82.7	—	Close	Near Claude Opus 4.8
MCP-Atlas (tool use)	76.8 / 77.0	75.3	—	GLM 5.2 leads
Tool-Decathlon	48.2	—	Higher	Claude leads
HLE (with tools)	40.5 / 54.7	52.2	—	GLM 5.2 leads
GPQA Diamond	91.2	—	—	Strong reasoning
AIME 2026	99.2	—	—	Strong math
Design Arena ELO	1360 (#1)	—	—	Ahead of Claude Fable 5
Code Arena (frontend)	#2 overall	—	—	Behind Claude Fable 5 only
FrontierSWE	#3 overall	—	—	Behind Fable 5 and Opus 4.8
Agent Arena (Max)	#10 overall, #1 open	—	—	Top open-weight agent

*Z.ai published no benchmarks at launch on June 13, 2026. The figures below come from the official benchmark table released alongside the open weights on June 16–17, plus independent third-party evaluations where noted.

Where GLM 5.2 wins: SWE-bench Pro is probably the headline number. 62.1 vs GPT-5.5's 58.6 on one of the more rigorous real-world coding benchmarks available. It also leads on MCP-Atlas (tool use), which matters specifically for agentic workflows. GPQA Diamond at 91.2 and AIME 2026 at 99.2 show solid reasoning capability beyond just code.

Where it doesn't win: Claude Opus 4.8 leads on Tool-Decathlon, and both Claude Fable 5 and Opus 4.8 sit ahead on FrontierSWE. The Design Arena #1 ranking is a genuine surprise for a model positioned primarily around coding.

The Agent Arena result deserves a separate mention: #10 overall but #1 among open-weight models. For teams who want the option to self-host, that's the number that matters most.

What's New in GLM 5.2

Dual thinking-effort levels

High: Balances capability and latency for everyday coding tasks
Max: Engages deeper reasoning for harder problems, equivalent to xhigh on some providers
Gives developers direct control over cost-vs-speed at inference time

IndexShare architecture

Covered above. The 2.9x FLOPs reduction at 1M context is the headline efficiency gain
Paired with an improved MTP layer for up to 20% faster speculative decoding

GLM Coding Plan access

Z.ai launched GLM 5.2 through its Coding Plan (Lite / Pro / Max / Team) before opening the standalone API and open weights
Managed access for teams that don't want to handle API integration directly

Broad IDE and tool compatibility

GLM 5.2 uses an Anthropic-compatible endpoint
Works out of the box with Claude Code, Cline, Cursor, OpenCode, Roo Code, Goose, Crush, OpenClaw, and Kilo Code, requiring just a base URL and model swap, no custom client code

GLM 5.2 Pricing

GLM 5.2 is available through two paths: a managed subscription or a per-token API.

GLM Coding Plan (subscription)

Billed at the undiscounted monthly rate, the tiers are:

Lite: $18/month
Pro: $72/month
Max: $160/month
Team: seat-based pricing

Z.ai applies billing-cycle discounts: 10% off monthly billing, 20% off quarterly, and 30% off annual. On an annual plan, those same tiers drop to $12.60, $50.40, and $112 per month respectively. Verify current rates on z.ai before committing, as promotions shift frequently.

API pricing (per token)

OpenRouter / FriendliAI: $1.40 per million input tokens / $4.40 per million output tokens
Alternative OpenRouter tier: $1.20 input / $3.20 output
Self-hosted: No per-token cost on the MIT-licensed weights (compute cost only)

How that stacks up against the competition

Model	Input (per M tokens)	Output (per M tokens)	Combined
GLM 5.2	$1.40	$4.40	~$5.80
Claude Opus 4.8	$5.00	$25.00	~$30.00
GPT-5.5	$5.00	$30.00	~$35.00
GLM 5.2 (self-hosted)	—	—	Compute only

The cost story is straightforward: GLM 5.2 costs roughly 6x less than GPT-5.5 ($5.80 vs $35.00 combined per million tokens) and roughly 5x less than Claude Opus 4.8 ($5.80 vs $30.00), while matching or outperforming both on the coding benchmarks developers actually care about. At high volumes, that gap compounds fast.

GLM 5.1 vs GLM 5.2: What Actually Changed

If you're already running GLM 5.1, which introduced the Coding Plan, here's what the upgrade gets you:

SWE-bench: Moves from ~58.4 (Verified) in 5.1 to 62.1 (Pro) in 5.2 (a harder benchmark, a better score)
IndexShare: New in 5.2; cuts per-token FLOPs by 2.9x at 1M context
Improved MTP layer: New in 5.2; raises speculative decoding throughput by up to 20%
Dual thinking-effort levels: High and Max didn't exist in 5.1
MIT license and open weights: Unchanged; same approach, same flexibility

It's not a ground-up rebuild. It's a targeted set of improvements on top of a model that was already competitive.

Regional Availability: What the Export Controls Mean

On June 12, 2026, the US Commerce Department issued an export control directive restricting access to Claude Fable 5 and Claude Mythos 5 for developers outside the United States.

For teams in affected regions, the shortlist of capable frontier coding models has narrowed. GLM 5.2's MIT license carries no regional access restrictions, making it the most capable open-licensed coding model currently available by benchmark ranking for developers who've lost access to Anthropic's latest models.

Evaluating GLM 5.2 for Agentic Workflows

Teams building agentic data workflows or analytics agents on tools like Agent Builder often face a real tradeoff: context length requirements vs. per-session cost.

GLM 5.2's 1M-token window and ~$5.80 combined per-million-token cost make it a practical option for high-volume, long-session agentic tasks where context accumulates across an entire codebase or dataset.

Start Building With Agent Builder

FAQs: GLM 5.2

What is GLM 5.2?

GLM 5.2 is a 753B parameter Mixture-of-Experts model from Z.ai, built for coding and agentic tasks. It features a 1M-token context window, MIT open-source license, and two thinking-effort modes: High and Max. It was officially announced on June 13, 2026, with open weights released on June 16–17, 2026.

Who made GLM 5.2?

Z.ai made GLM 5.2. Z.ai is the international brand of Zhipu AI, a company spun out of Tsinghua University that has developed the GLM model family across multiple generations.

When was GLM 5.2 released?

Z.ai announced GLM 5.2 on June 13, 2026. Open weights became available on HuggingFace on June 16–17, 2026, alongside the official benchmark table. No benchmark data was published at the initial launch.

What is the GLM 5.2 context window?

GLM 5.2 supports a 1,000,000-token context window via the glm-5.2[1m] model ID, with up to 131,072 output tokens supported.

What are GLM 5.2's benchmark scores?

GLM 5.2 scores 62.1 on SWE-bench Pro, 76.8–77.0 on MCP-Atlas, 91.2 on GPQA Diamond, and 99.2 on AIME 2026. It ranks #1 on Design Arena ELO and #1 among open-weight models on Agent Arena in Max mode. Full figures are in the benchmark table above.

How does GLM 5.2 compare to Claude Opus 4.8?

GLM 5.2 leads Claude Opus 4.8 on SWE-bench Pro and MCP-Atlas. Claude Opus 4.8 leads on Tool-Decathlon and FrontierSWE. On price, GLM 5.2 comes in at roughly $5.80 per million combined tokens via API; Claude Opus 4.8 runs about $30.00, roughly 5x more expensive.

How much does GLM 5.2 cost?

Via API, GLM 5.2 costs $1.40 per million input tokens and $4.40 per million output tokens through OpenRouter or FriendliAI. The GLM Coding Plan undiscounted monthly rates are $18 for Lite, $72 for Pro, and $160 for Max. Annual billing drops those to $12.60, $50.40, and $112 per month. Self-hosting on the MIT-licensed weights costs nothing per token.

Is GLM 5.2 open source?

Yes. GLM 5.2 is MIT-licensed. Weights are on HuggingFace at zai-org/GLM-5.2 for the standard model and zai-org/GLM-5.2-FP8 for the quantized, lower-memory-footprint version, with no regional restrictions on use or distribution.

What is IndexShare?

IndexShare is an architectural feature in GLM 5.2 that reuses the same attention indexer across every four sparse attention layers, instead of computing a separate index per layer. It cuts per-token FLOPs by 2.9x at 1M context length, making long-context inference significantly cheaper to run.

Related articles

May 2026 Security Incident

Why We're Open Sourcing OpenUI

AI Copilots: A Complete Guide to Intelligent Assistance

What Are AI Agents: Types, Capabilities, and Real-World Use

ChatKit: What It Is, What It Isn’t, and Where C1 Fits In

PowerPoint Automation Reimagined with C1 Artifacts API

The Future of AI Interfaces is Generative

MCP UI: Making AI Apps Interactive

What Is AG-UI and Why It Matters for AI-Native User Interfaces

How to design AI-Native Conversational Interfaces : From Templates to Generative UI

GPT 5 vs. GPT 4.1

How to Build Generative UI Applications

Implementing Generative Analytics with Thesys and MCP

Evolution of Analytics: From Static Dashboards to Generative UI

Why Generating Code for Generative UI is a bad idea

Building the First Generative UI API: Technical Architecture and Design Decisions Behind C1

How we evaluate LLMs for Generative UI

Generative UI vs Prompt to UI vs Prompt to Design

Generative UI: What It Is and Why It Matters