# GLM 5.2 Tops Open Models at 51 on AA Index

> Source: [https://icharles.com/articles/glm-52-open-model-benchmark-export-ban](https://icharles.com/articles/glm-52-open-model-benchmark-export-ban) (canonical)
> Author: iCharles News — iCharles, https://icharles.com
> Published: 2026-06-29

## TL;DR

Zhipu AI released GLM 5.2 on June 13, 2026 — a 744-billion-parameter open-weight model that scores 51 on the Artificial Analysis Intelligence Index v4.1, the highest of any open-weight model and ahead of Google's Gemini 3.1 Pro Preview at 46. It costs $1.40 per million input tokens via OpenRouter, roughly one-sixth the price of GPT-5.5. On the same day, the US government ordered Anthropic to suspend foreign access to its most advanced models, Fable 5 and Mythos 5.

## What is GLM 5.2 and what did it score?

**GLM 5.2** is an open-weight mixture-of-experts language model from Chinese AI company Zhipu AI, released on June 13, 2026. It scores 51 on the Artificial Analysis Intelligence Index v4.1 — the highest score among all open-weight models. That puts it ahead of Google's Gemini 3.1 Pro Preview (46) and Gemini 3.5 Flash (50), [according to MLQ.ai](https://mlq.ai/news/zhipus-glm-52-surges-past-google-models-as-us-export-ban-hobbles-anthropic/).

The model has 744 billion total parameters and 40 billion active parameters per token. Its context window is one million tokens — four times larger than its predecessor. Maximum output per response is 131,072 tokens.

## How does GLM 5.2 cut compute costs?

GLM 5.2 introduces a technique called **IndexShare**. It reuses the same indexer across every four sparse attention layers. Zhipu says this cuts per-token compute by about 2.9 times at full context length.

Via OpenRouter, the model costs $1.40 per million input tokens and $4.40 per million output tokens. That compares to $5/$30 for GPT-5.5 and $5/$25 for Claude Opus.

## How does GLM 5.2 compare to other models on benchmarks?

| Benchmark | GLM 5.2 | Claude Opus 4.8 | Gemini 3.1 Pro Preview |
|---|---|---|---|
| AA Intelligence Index v4.1 | 51 | — | 46 |
| SWE-bench Pro | 62.1 | 69.2 | — |
| LMArena Code Arena rank | 2nd (1,595 pts) | — | — |
| LMArena Design Arena rank | 1st | — | — |
| GDPval-AA v2 | ~1,524 | — | — |

On SWE-bench Pro, GLM 5.2 scores 62.1 versus Claude Opus 4.8's 69.2 — a seven-point gap that narrows to roughly one percentage point on long-horizon coding tasks, [per MLQ.ai](https://mlq.ai/news/zhipus-glm-52-surges-past-google-models-as-us-export-ban-hobbles-anthropic/).

## How did GLM 5.2 perform on the CyBT-CTF cybersecurity benchmark?

Researchers at Louie.ai tested GLM 5.2 on the **CyBT-CTF**, a cheating-resistant agentic cybersecurity investigation benchmark. GLM 5.2 achieved a 28/59 solve rate — matching Claude Opus 4.7 and 4.8 on the same benchmark, [according to Graphistry](https://www.graphistry.com/blog/glm-5-2-cybersecurity-open-model).

The next-best open model, MiniMax 2.5, scored only 16/59. Western open model GPT-open-120B scored 12/59. Graphistry noted that Claude Opus 4.8 runs 19% faster than OpenCode/GLM 5.2, but costs 2.2 times more for the same results.

Graphistry also flagged a concern: GLM 5.2's correct and incorrect answers are highly correlated with both GPT-5.5 and Opus 4.8. Cohen's Kappa scores of 0.80 and 0.76 respectively — compared to 0.63 between OpenAI and Anthropic — led the researchers to ask whether GLM 5.2 may be the result of model distillation from frontier providers. Anthropic had previously reported that Chinese-origin model companies were attempting distillation attacks, though Zhipu was not named in that report.

Here's what we know so far: the distillation question remains open, and Graphistry has flagged it as a warning, not a confirmed finding.

## What US export control action hit Anthropic on the same day?

The US government issued an export-control directive on June 12, 2026. It ordered Anthropic to disable access to **Fable 5** and **Mythos 5** — its two most advanced models — for all foreign nationals. The cutoff took effect at 5:21 p.m. ET.

The directive cited a discovered jailbreak method that bypassed safeguards on Mythos's advanced cybersecurity capabilities. Anthropic complied but pushed back publicly. The company stated the jailbreak was narrow and that "the same jailbreak could be used to elicit similar capabilities from other publicly available models, including OpenAI's GPT-5.5, that are not subject to similar national security export controls." Access to all other Anthropic models remained unaffected.

This situation connects to broader US government AI policy — including [Trump's EO on AI model access](/articles/trump-eo-ai-model-access) — and raises questions about how export controls apply unevenly across frontier providers. It also adds context to [Anthropic's salary disclosures](/articles/anthropic-salary-h1b-filings-2026), which showed heavy reliance on foreign-national talent that would now be cut off from the company's top models.

Zhipu released GLM 5.2 under MIT license on the same day. The company stated: "frontier intelligence should not belong to only a few people, nor be subject to withdrawal by a handful of rules at any moment." The model ships with native support for over 20 agent tools, including Claude Code, Cline, and Cursor.

## How did markets react to GLM 5.2?

**Knowledge Atlas Technology** (HKEX: 2513), Zhipu's Hong Kong-listed parent company, surged as much as 48% on the announcement day. The stock closed up 32.8% at HK$1,457. It later reached an all-time high of HK$2,980 on June 22 before pulling back.

The stock has gained roughly 820% since its January 2026 IPO. Its market cap now exceeds HK$1 trillion, or about $130 billion.

- JPMorgan raised its price target on Knowledge Atlas from HK$950 to HK$1,400
- Macquarie maintained a Buy rating
- CLSA issued a Hold rating on June 23

The stock's rise follows a broader pattern of investor attention to open-source AI infrastructure — similar to moves seen around [OpenAI's GPT-5.6 rollout](/articles/openai-gpt-56-trusted-partners-release) and [Oracle's AI-driven restructuring](/articles/oracle-layoffs-21000-ai-expansion).

## Where can developers access GLM 5.2?

The model weights are available on Hugging Face and ModelScope under MIT license. Developers can self-host, fine-tune, and use GLM 5.2 commercially with no restrictions and no regional locks. API access is available via OpenRouter at $1.40 per million input tokens.

<!--il:v1-->

## Frequently asked questions

**What score did GLM 5.2 get on the Artificial Analysis Intelligence Index?**

GLM 5.2 scored 51 on the Artificial Analysis Intelligence Index v4.1, making it the highest-scoring open-weight model on that benchmark. It outscored Google's Gemini 3.1 Pro Preview, which scored 46, and Gemini 3.5 Flash, which scored 50. The score puts GLM 5.2 near parity with GPT-5.5 on long-horizon coding benchmarks, according to MLQ.ai.

**How much does GLM 5.2 cost compared to GPT-5.5 and Claude Opus?**

Via OpenRouter, GLM 5.2 costs $1.40 per million input tokens and $4.40 per million output tokens. GPT-5.5 costs $5 per million input tokens and $30 per million output tokens. Claude Opus costs $5 input and $25 output. That makes GLM 5.2 roughly one-sixth the input price of its closed-source competitors.

**Why did Anthropic suspend Fable 5 and Mythos 5 access?**

The US government issued an export-control directive on June 12, 2026, ordering Anthropic to suspend access to Fable 5 and Mythos 5 for all foreign nationals. The directive cited a discovered jailbreak that bypassed safeguards on Mythos's advanced cybersecurity capabilities. Anthropic complied but stated it believed the jailbreak was narrow and that similar techniques could affect other models not subject to the same controls.

**What is Knowledge Atlas Technology and how did its stock perform?**

Knowledge Atlas Technology (HKEX: 2513) is the Hong Kong-listed parent company of Zhipu AI. Its stock surged as much as 48% on the day GLM 5.2 was announced, closing up 32.8% at HK$1,457. The stock hit an all-time high of HK$2,980 on June 22, 2026. Since its January 2026 IPO, shares have gained roughly 820%, giving the company a market cap above HK$1 trillion.

**Did GLM 5.2 match Anthropic's models on cybersecurity benchmarks?**

On the CyBT-CTF agentic cybersecurity benchmark, GLM 5.2 achieved a 28/59 solve rate, matching Claude Opus 4.7 and 4.8. The next-best open model, MiniMax 2.5, scored only 16/59. Graphistry noted that Opus 4.8 runs 19% faster but costs 2.2 times more. Graphistry also flagged that GLM 5.2's answer patterns are highly correlated with both GPT-5.5 and Opus 4.8, raising questions about potential model distillation.