What is GLM 5.2 and who made it?

GLM 5.2 is a 744-billion-parameter Mixture-of-Experts large language model released by Z.ai, the international brand of Zhipu AI, a Beijing lab founded out of Tsinghua University. It launched on June 13, 2026, under a permissive MIT open-source license with weights available on Hugging Face. The model is specifically designed for long-horizon agentic coding and software engineering tasks.

How does GLM 5.2 compare to Claude Opus 4.8 on coding benchmarks?

On FrontierSWE, a benchmark for multi-hour technical projects, GLM 5.2 scores 74.4 percent versus Claude Opus 4.8's 75.1 percent, a gap of less than one percentage point. On Terminal-Bench 2.1, GLM 5.2 scores 81.0, close to but behind Opus 4.8 (85.0). On the Code Arena Frontend human-preference leaderboard, GLM 5.2 ranks second overall ahead of both Claude Opus 4.7 and 4.8 Thinking modes. That said, Claude Opus 4.8 retains meaningful leads on the hardest long-horizon tasks, including SWE-Marathon and NL2Repo.

How much cheaper is GLM 5.2 compared to proprietary frontier models?

As of June 2026, GLM 5.2's API is priced at $1.40 per million input tokens and $4.40 per million output tokens. Claude Opus 4.8 is priced at $5 per million input and $25 per million output. That makes GLM 5.2 roughly 3.6 times cheaper on input and 5.7 times cheaper on output. For organizations self-hosting the open weights under the MIT license, the per-token cost drops to zero (beyond infrastructure costs).

What is the IndexShare innovation and why does it matter?

IndexShare is an attention optimization that Z.ai introduced in GLM 5.2. In standard sparse attention, every transformer layer runs its own indexer to identify which tokens to attend to. IndexShare instead runs the full indexer only once every four layers and shares those selected token indices across the following three layers. Z.ai reports this cuts per-token FLOPs by 2.9 times at a one-million-token context length, which is what makes the 1M context window practically affordable to serve.

Can GLM 5.2 run locally, and what hardware does it need?

Yes, with caveats. The full-precision model requires approximately 8 H200 GPUs for production serving. Quantized versions reduce this significantly: INT4 quantization fits on roughly 4 H200 GPUs, and Unsloth's 1-bit GGUF quantization runs at around 21.6 tokens per second on a Mac Studio M3 Ultra with 256GB unified memory. The MIT license permits anyone to download, run, and fine-tune the model without usage restrictions, but the hardware barrier is real and should not be understated.

Does the open-weight status of GLM 5.2 actually matter beyond cost?

Yes, significantly. An MIT-licensed open-weight model can be fine-tuned on proprietary codebases, run in air-gapped environments with no external API dependency, pinned to a specific version indefinitely, and deployed anywhere without regional restrictions. For organizations in regulated industries, defense-adjacent contexts, or those with strict data-residency requirements, these properties can be more important than benchmark scores. The open-weight format also allows distillation into smaller models for edge deployment, something impossible with closed API-only systems.

AI Models & Releases

GLM 5.2: How a 744B Open Model Is Closing the Gap With Claude Opus on Code

By The Agent5 founder·June 29, 2026·How we fact-check

Z.ai's GLM 5.2 is a 744-billion-parameter open-weight model that scores within one percentage point of Claude Opus 4.8 on frontier coding benchmarks, under an MIT license, at roughly one-sixth the API cost. That combination is reshaping how teams think about model pricing and the open-versus-closed divide.

ShareX LinkedIn WhatsApp

Key takeaways

GLM 5.2 is a 744B open-weight MoE model from Z.ai that scores within one percentage point of Claude Opus 4.8 on FrontierSWE, the benchmark for multi-hour agentic coding projects, while costing 5.7 times less per output token via its API.
The model's IndexShare architecture reduces per-token computation by 2.9 times at a one-million-token context length, making the full 1M context window practically affordable to serve rather than a theoretical maximum.
An MIT license with no regional restrictions means teams can fine-tune, self-host, air-gap, and version-pin the model, advantages that are independent of benchmark scores and often decisive for regulated or security-sensitive use cases.
Claude Opus 4.8 retains clear leads on the hardest long-horizon tasks (SWE-Marathon, NL2Repo), so GLM 5.2 is not a universal replacement; the right prediction is near-parity on mainstream agentic coding, not across-the-board superiority.
The open-versus-closed performance gap is closing generation over generation, and the pricing pressure this creates on proprietary frontier models is likely to intensify as the trend continues into future releases.

GLM 5.2: How a 744B Open Model Is Closing the Gap With Claude Opus on Code

What GLM 5.2 Actually Is

The Architecture Trick That Makes 1 Million Tokens Affordable

The Benchmark Picture: Close, But Not a Sweep

What the Pricing Gap Actually Means

A Closer Look at the Open-Weight Advantage

Where It Falls Short and What to Watch

The Geopolitical Undercurrent

What This Means for How You Think About Model Pricing

Sources