benchmark

Zhipu's GLM-5.2 matches Anthropic's Claude Opus 4.8 on agentic benchmark at one-fifth the cost

TL;DR

Zhipu AI's open-source GLM-5.2 model scores within one percentage point of Anthropic's Claude Opus 4.8 on a key agentic benchmark while costing approximately one-fifth as much. The release comes as U.S. government restrictions limit access to Anthropic's Fable and OpenAI's GPT-5.6 models.

2 min read
0

GLM-5.2 at a glance

Context window1049K tokens
Input$0.62/1M tokens
Output$2.42/1M tokens

How GLM-5.2 ranks now

BenchmarkScoreRankSource
Arena Elo1475 Elo27 of 141LMArena, Nov 2026 ↗
GPQA85.6%83 of 165Vals AI, Sep 2026 ↗
SWE-bench Verified82.8%10 of 80Vals AI, Sep 2026 ↗
Speed (tok/s)62 tok/s74 of 105Artificial Analysis, Nov 2026 ↗
AIME 202699.2%1 of 26llm-stats, Oct 2026 ↗

Rank is against every model in our database with a sourced score on that benchmark, each counted once with its best score.

All GLM-5.2 specs and scores →

Chinese open-source model challenges frontier labs on price-performance

Zhipu AI's GLM-5.2 scores within one percentage point of Anthropic's Claude Opus 4.8 on agentic benchmarks while costing roughly 20% as much, according to the Chinese AI startup. The open-source model, released last week, has surpassed all other open releases on the closely watched benchmark.

OpenRouter token traffic for GLM-5.2 is climbing faster than it did following DeepSeek's V4 launch in April, suggesting rapid developer adoption. Unlike DeepSeek, which focused primarily on chat applications, GLM-5.2 demonstrates strength in agentic tasks including planning, coding, testing, and task looping—capabilities enterprises are prioritizing for automation.

Enterprise cost pressures drive adoption

"I've been consistently surprised by how quickly the open source has caught up," Gabe Pereyra, co-founder of legal AI company Harvey, told CNBC. "GLM-5.2, you're seeing the first model where it's really competitive with some of these closed-source frontier models."

As token spend strains enterprise AI budgets, "intelligence per dollar" is emerging as the critical metric. GLM-5.2's combination of competitive performance and significantly lower costs addresses this pressure directly. The model is free to download, fine-tune, and deploy on enterprise infrastructure, eliminating recurring API costs.

U.S. restrictions create opening for open source

The timing coincides with increased U.S. government oversight of frontier models. Anthropic pulled its Fable Mythos-class model following a Trump administration order. OpenAI announced Friday it is limiting access to GPT-5.6 models "at the request of the U.S. government."

These restrictions make models that "no one can revoke" more attractive to enterprises concerned about deployment stability, according to the report. Open-source models eliminate dependency on external API access and regulatory compliance from third-party providers.

Specific benchmark numbers not disclosed

While the article states GLM-5.2 lands "within a percentage point" of Claude Opus 4.8 on "a key agentic benchmark," neither the specific benchmark name nor exact scores are disclosed. The pricing comparison—"roughly a fifth of the cost"—also lacks precise per-token figures for verification.

Zhipu AI, based in Beijing, has been building foundation models since 2019. The company previously released GLM-4 and other models in its series.

What this means

If verified, GLM-5.2's performance represents a significant compression in the gap between open-source and frontier closed models, particularly for agentic workflows. The combination of regulatory uncertainty around U.S. models and budget pressures on token spend could accelerate enterprise adoption of open-source alternatives. However, the lack of disclosed benchmark specifics makes it difficult to independently verify the claimed parity with Claude Opus 4.8. The shift toward "intelligence per dollar" as the primary evaluation metric reflects a maturing market moving beyond pure capability races.

Source: cnbc.com ↗

Related Articles

benchmark

AWS launches SageMaker JumpStart Model Leaderboard covering more than 70 models

Amazon Web Services introduced the Amazon SageMaker JumpStart Model Leaderboard on November 18, 2026. It gives SageMaker Studio users one view of benchmark results, response performance and estimated instance pricing across more than 70 models.

benchmark

Vals tests: Claude Opus 5.5 picked Minab school as military target in 22 of 32 runs, Bloomberg reports

AI models from eight companies often marked civilians, ambulances, surrendering soldiers and a school in Iran as military targets in simulated tests, Bloomberg reported, citing a report by evaluator Vals. Anthropic's models did worst, while OpenAI's were among the best but still made mistakes.

benchmark

Claude Fable 5.2 paints images through Python code, using about 700,000 simulated strokes in one piece

Anthropic's Claude Fable 5.2 produces images by writing Python programs that drive a simulated paint engine, with no external image model. Anthropic reports 64.9 percent on its internal RoboDojo Sim benchmark against 29.8 percent for OpenAI's GPT-6 Astra, though a real-world robot arm test showed limits in speed and fine motor control.

benchmark

Latent Space test: NaiveAI's Naive-N0.5-Flash hits about 2,100 tokens/s, 3.1s median response

Latent Space tested NaiveAI's open-source Naive-N0.5-Flash against DeepSeek V4.1 Flash 1017 and GLM 5.3 Flash. It measured a peak of about 2,100 tokens/s and a 3.1-second median response time with thinking on, and the model scored 71.9 on NL2Repo-Bench.

Comments

Loading...