GLM-5.2 open weights beat GPT-5.5 at coding

What happened

On 13 June 2026 Z.ai (formerly Zhipu AI) published the full weights of GLM-5.2 on Hugging Face (zai-org/GLM-5.2) and ModelScope under the MIT license — no regional restrictions, with the right to self-host, fine-tune and use commercially.

The essentials:

• Architecture: MoE with ~753B parameters, ~40B active per token; a new sparse-attention technique called IndexShare keeps inference cost in check at a 1M-token context.
• Benchmarks: SWE-bench Pro — 62.1 versus 58.6 for GPT-5.5 and 58.4 for GLM-5.1; Artificial Analysis Intelligence Index v4.1 — 51 points, ahead of MiniMax-M3 and DeepSeek V4 Pro (VentureBeat).
• API pricing: $1.40/$4.40 per million tokens — roughly one sixth the cost of comparable frontier models.
• Release context: the weights landed a day after the US Commerce Department ordered Anthropic to cut off Fable 5 and Mythos 5 for foreign users, which makes the timing look like an answer to export restrictions (TechTimes).

Why it matters

• Open weights have caught up with the frontier on coding. For the first time an MIT-licensed model beats a closed GPT-5.5 on long-horizon engineering tasks — self-hosting coding agents becomes a real alternative.
• Price pressure. At 1/6th of frontier cost it reinforces the trend set by Grok 4.5 and Muse Spark: paying for a top-tier model is getting harder to justify.
• The geopolitics of open source. A release timed against export controls shows that restrictions accelerate rather than slow the spread of open models. That said, using the Z.ai API carries the risk of sending data into Chinese jurisdiction — for sensitive code, pick self-hosting.