GLM-5.2: Open-Weight Model Beats GPT-5.5 on Coding at One-Sixth the Cost

Z.ai — the international arm of Chinese AI lab Zhipu AI — released GLM-5.2 under a permissive MIT license, and it outperforms GPT-5.5 on long-horizon coding benchmarks at approximately one-sixth the API cost.

The numbers are concrete. On SWE-Bench Pro, the industry benchmark for real-world software engineering tasks, GLM-5.2 scores 62.1%. GPT-5.5 scores 58.6%. GLM-5.2 also holds the top spot on AIME 2026 at 99.2%. API pricing: $1.40 per million input tokens, $4.40 per million output tokens. GPT-5.5 is $5.00 input, $30.00 output. For teams running AI coding agents at scale, that gap is not marginal — it is a structural cost advantage that compounds across every inference call.

The architecture is a 753-billion-parameter Mixture-of-Experts model with 40 billion active parameters per forward pass and a one-million-token context window. The key technical innovation is IndexShare, a sparse attention mechanism that reduces per-token FLOPs by 2.9× at million-token context lengths. At that scale, most models become computationally prohibitive. GLM-5.2 makes 1M-context economically viable for routine use, not just for one-off tasks.

The MIT license means no commercial restrictions. Any company, anywhere, can run GLM-5.2 internally, fine-tune it, build products on top of it, or host it for others — without licensing fees or usage reports to Z.ai. For the open-source ecosystem, that is the headline: a frontier-class coding model that is now genuinely free.

The geopolitical framing is intentional. Zhipu built GLM-5.2 on domestic hardware under U.S. chip export control restrictions that limit Chinese labs’ access to NVIDIA’s most advanced accelerators. That it still outperforms on key benchmarks makes the point Z.ai is trying to make — that export controls slow but do not stop frontier development.

Cloudflare moved quickly. GLM-5.2 was added to Workers AI on June 16, accessible via REST API at model ID @cf/zai-org/glm-5.2 with a 262,144-token initial context window and a 1M-token window planned. Zhipu’s Hong Kong-listed stock surged approximately 33% on the launch day.

The open-weight coding leaderboard in June 2026 now reads: GLM-5.2 at 62.1%, Kimi K2.7-Code at 62.0%, MiniMax M3 at 59.0%. All three are open-weight. All three are cheaper than the closed-weight alternatives by a factor of 4–6×. The argument that frontier AI requires closed systems is getting harder to sustain.

Sources

• VentureBeat
• TechTimes