# GLM-5.2

Z.ai, released 2026-06-13

**Verdict: conditional.** Top open-weight coding model with 1M-token context and MIT license. It beats GPT-5.5 on multiple benchmarks at a fraction of cost, but still trails Claude Opus 4.8 on the hardest long-horizon tasks. Last verified 2026-06-23.

## Pricing

- Input: $1.40 / 1M tokens
- Output: $4.40 / 1M tokens
- Context: 1M tokens

## Strengths

- Strongest open-weight coding model at launch — 62.1 on SWE-bench Pro, 81.0 on Terminal-Bench 2.1
- 1M-token context window — 5× larger than GLM-5.1; can process entire codebases in one prompt
- MIT license with open weights — fully self-hostable, no revenue restrictions, commercial use allowed
- Massive improvement over GLM-5.1 — Terminal-Bench 2.1 jumped from 62.0 to 81.0; DeepSWE from 18.0 to 46.2
- IndexShare architecture cuts per-token FLOPs by 2.9× at 1M context — makes long-context coding economically viable
- Dual reasoning-effort modes (High and Max) — trade speed for depth
- Heavy token appetite — some reports of high token consumption for equivalent tasks vs direct API calls
- Training code and full technical report not released at launch — only inference weights under MIT
- Still trails Claude Opus 4.8 on hardest tasks (SWE-bench Pro: 62.1 vs 69.2; SWE-Marathon: 13.0 vs 26.0)

## Benchmarks

- SWE-bench Pro: 62.1 (https://www.labellerr.com/blog/glm-5-2-open-weight-ai-model/)
- Terminal-Bench 2.1 (Terminus-2): 81.0 (https://www.labellerr.com/blog/glm-5-2-open-weight-ai-model/)
- Terminal-Bench 2.1 (Best Harness, Claude Code): 82.7 (https://www.labellerr.com/blog/glm-5-2-open-weight-ai-model/)
- FrontierSWE: 74.4 (https://www.labellerr.com/blog/glm-5-2-open-weight-ai-model/)
- DeepSWE: 46.2 (https://www.labellerr.com/blog/glm-5-2-open-weight-ai-model/)
- AIME 2026: 99.2 (https://www.labellerr.com/blog/glm-5-2-open-weight-ai-model/)
- GPQA-Diamond: 91.2 (https://www.labellerr.com/blog/glm-5-2-open-weight-ai-model/)
- HLE (Humanity's Last Exam): 40.5 (https://www.labellerr.com/blog/glm-5-2-open-weight-ai-model/)
- HLE w/ Tools: 54.7 (https://www.labellerr.com/blog/glm-5-2-open-weight-ai-model/)
- SWE-Marathon: 13.0 (https://www.labellerr.com/blog/glm-5-2-open-weight-ai-model/)
- MCP-Atlas Public Set: 76.8 (https://www.labellerr.com/blog/glm-5-2-open-weight-ai-model/)
- PostTrainBench: 34.3 (https://www.labellerr.com/blog/glm-5-2-open-weight-ai-model/)

## Sources

- [Developer Tech](https://www.developer-tech.com/news/z-ai-glm-5-2-long-context-coding-agents/)
- [VentureBeat — GLM-5.2 Benchmark Coverage](https://venturebeat.com/technology/z-ais-open-weights-glm-5-2-beats-gpt-5-5-on-multiple-long-horizon-coding-benchmarks-for-1-6th-the-cost)
- [llm-stats.com — GLM-5.2 Pricing & Benchmarks](https://llm-stats.com/models/glm-5.2)
- [FelloAI — GLM-5.2 Complete Guide](https://felloai.com/glm-5-2/)
- [kilo.ai — GLM-5.2 Model Page](https://kilo.ai/models/z-ai-glm-5-2)
- [BuildFastWithAI — GLM-5.2 Review](https://www.buildfastwithai.com/blogs/glm-5-2-review-2026)
- [Labellerr — GLM-5.2 Beats GPT-5.5](https://www.labellerr.com/blog/glm-5-2-open-weight-ai-model/)
- [Codersera — GLM-5.2 Complete Guide](https://codersera.com/blog/glm-5-2-complete-guide-2026/)

Canonical: https://neomanex.com/models/glm-5-2