| Developer | Z.ai (formerly Zhipu AI) |
| Released | June 13, 2026 |
| License | MIT (open weights, commercial use allowed) |
| Intelligence | #1 open-weight — Artificial Analysis Intelligence Index |
| Speed | 102.3 output tokens/sec |
| Input price | $1.40 per 1M tokens |
| Output price | $4.40 per 1M tokens |
| Cache hit price | $0.26 per 1M tokens (81% discount) |
| Context window | 1M tokens (~750,000 words) |
| Total parameters | 753B |
| Active parameters | 40B (Mixture of Experts) |
Model Summary
GLM-5.2 is Z.ai’s open-weight reasoning model. It ranks #1 among open-weight models on the Artificial Analysis Intelligence Index and #4 overall out of 124 on BenchLM.ai’s provisional leaderboard. The model features a 1-million-token context window, dual reasoning modes (high and max effort), strong coding and math performance, and MIT licensing for unrestricted commercial use.
Compared to closed frontier models, GLM-5.2 is competitive on most coding benchmarks at roughly one-sixth the blended cost of GPT-5.5 or Claude Opus 4.8. Where it trails is on long-horizon planning tasks — Claude Opus 4.8 wins decisively on SWE-Marathon and similar sustained-reasoning benchmarks. For the 90% of AI tasks that don’t fall into that category, performance is comparable.
Technical Specifications
| Reasoning | Yes — high and max effort modes |
| Input modality | Text |
| Output modality | Text |
| Vision | Not supported (June 2026 release) |
| Context window | 1,000,000 tokens |
| Max output tokens | 131,072 |
| Architecture | Mixture of Experts (MoE) |
| Total parameters | 753B |
| Active parameters | 40B per forward pass |
| Checkpoints | BF16 and native FP8 |
| License | MIT |
What Changed from GLM-5.1
- Context window: 200K to 1M tokens (5x increase)
- Max output: 120K to 131,072 tokens per response
- Reasoning modes: New dual-mode setup — high and max effort
- IndexShare: Reuses sparse-attention top-k indices across every 4 layers, cutting FLOPs at 1M context by 2.9x vs GLM-5.1
- Multi-Token Prediction: Extended from 3 to 5 draft tokens per step (faster throughput on coding and agentic tasks)
- Native FP8 checkpoint: Added alongside BF16 for reduced memory requirements
Benchmarks
Coding
| Benchmark | GLM-5.2 | Claude Opus 4.8 | GPT-5.5 |
|---|---|---|---|
| SWE-bench Pro | 62.1 | — | 58.6 |
| FrontierSWE | 74.4 | 75.1 | — |
| MCP-Atlas | 76.8 | 77.8 | — |
| SWE-Marathon | — | Wins by ~2x | — |
Math and Rankings
- AIME 2026: 99.2%
- Design Arena Code Categories: #1 overall (beats Claude Fable 5 by 10 Elo)
- Artificial Analysis Intelligence Index: #1 open-weight at launch
- BenchLM.ai overall: #4 out of 124 (provisional), #9 out of 33 (verified)
Pricing
| GLM-5.2 | Claude Opus 4.8 | GPT-5.5 | |
|---|---|---|---|
| Input (per 1M tokens) | $1.40 | $5.00 | $5.00 |
| Output (per 1M tokens) | $4.40 | $25.00 | $30.00 |
| Cache hit (per 1M) | $0.26 | — | — |
Subscription plans: Lite ~$3–6/month, Pro ~$15–19/month, Max ~$80/month. Free tier available for API testing.
How to Access GLM-5.2
Direct API (Z.ai): Launched June 16, 2026. Uses an Anthropic-compatible endpoint. Drop-in replacement for Claude Code, Cline, Roo Code, Goose, and OpenCode with a base URL swap and model name change. No framework changes required.
Third-party providers: Available on Together AI, OpenRouter, Requesty, WaveSpeed, and others.
Local inference: Supported on vLLM, SGLang, Transformers, and KTransformers. Available on Ollama. Full 1M-context inference requires multiple high-memory GPUs. Quantized versions are more practical for most local setups.
Who It’s For
Developers and AI engineers doing agentic coding, multi-file code review, or whole-repo navigation. SWE-bench and code arena scores are among the best available at any price point.
Researchers and analysts working with large document sets. The 1M context window handles full corpora without chunking or multi-pass retrieval on most realistic document sizes.
High-volume workflow builders who have been paying frontier rates for structured generation, data extraction, or code tasks. At one-sixth the blended cost with comparable performance, the savings compound quickly at scale.
Self-hosting teams that need open weights for compliance, latency, or infrastructure reasons. MIT license, no usage restrictions, runs on standard inference stacks.
Verdict
GLM-5.2 is the most significant open-weight model release of 2026 so far. It competes with frontier closed models on coding and math, costs roughly one-sixth as much, and ships under an MIT license with no usage restrictions.
For most coding and structured generation tasks, performance is comparable to GPT-5.5 and within a few points of Claude Opus 4.8. Route complex multi-step planning and long-horizon engineering to Opus 4.8. Route the rest here.
Frequently Asked Questions
What does GLM stand for?
GLM stands for General Language Model. It’s the model family developed by Z.ai (formerly Zhipu AI), a Chinese AI research company spun out of Tsinghua University.
Is GLM-5.2 open source?
Yes. Released under an MIT license, making it freely usable, modifiable, and deployable for commercial purposes.
Is GLM-5.2 free to use?
Z.ai offers a free tier for API testing. Pay-as-you-go: $1.40 per 1M input tokens, $4.40 per 1M output tokens. Subscriptions from ~$3/month.
How does GLM-5.2 compare to GPT-5.5?
Competitive on coding at one-sixth the cost. GLM-5.2 beats GPT-5.5 on SWE-bench Pro (62.1 vs 58.6) and leads on design and code arenas. GPT-5.5 holds an edge on some general instruction-following tasks.
How does GLM-5.2 compare to Claude Opus 4.8?
Within 1–2 points on FrontierSWE and MCP-Atlas. Opus 4.8 wins decisively on SWE-Marathon and long-horizon planning. GLM-5.2 leads on code arenas and costs significantly less.
Can GLM-5.2 run locally?
Yes. Runs on vLLM, SGLang, Transformers, and KTransformers. Available on Ollama. Full 1M-context inference requires multiple high-memory GPUs; quantized versions are more practical for most local setups.
What is the GLM-5.2 context window?
1 million tokens (~750,000 words). The IndexShare architecture makes this a usable window rather than a theoretical ceiling, by reducing per-token FLOPs at long range by 2.9x vs the previous version.
Does GLM-5.2 support vision?
No. Vision support is not included in the June 2026 release. Text and code only.
Is GLM-5.2 a reasoning model?
Yes. It supports high and max effort reasoning modes. Max effort runs longer chain-of-thought traces for harder problems at the cost of higher latency and token usage.
What is the license for GLM-5.2?
MIT license. Commercial use allowed. Weights are publicly available for download and self-hosting.
How many parameters does GLM-5.2 have?
753 billion total parameters. 40 billion active per forward pass. It’s a Mixture of Experts model, so only the active parameters run computation on each token.