Z.ai released GLM-5.2 on June 13, 2026, making it immediately available to every GLM Coding Plan tier — Lite, Pro, Max, and Team — before publishing any benchmark scores or opening the standalone API. The model ships with a 1-million-token context window, two configurable thinking-effort levels (High and Max), and MIT-licensed open weights that arrived on Hugging Face three days later. For developers, the practical signal is clear: GLM-5.2 is a production-ready coding agent today, while GLM-5.5 remains entirely unconfirmed.

The launch sequence was unusual. Z.ai prioritized distribution through the Coding Plan over the typical benchmark-first rollout. API and chatbot access followed on June 16 alongside the open-weight release and a full benchmark scorecard showing GLM-5.2 reaching 81.0 on Terminal-Bench 2.1 and 62.1 on SWE-bench Pro — within striking distance of Claude Opus 4.8. Independent evaluators on Arena's Code Arena Frontend board ranked it #2 among open models, ahead of Opus 4.7 and 4.8 in thinking mode. The model also supports eight coding agents including Claude Code, Cline, Roo Code, Goose, and OpenCode via OpenAI-compatible and Anthropic Messages APIs, requiring only an environment-variable change to switch.

Confirmed

  • Model identity: GLM-5.2 is a 744B-parameter Mixture-of-Experts model with 40B active parameters per token, built on the GLM-5 backbone with IndexShare sparse attention (2.9x FLOPs reduction at 1M context) and improved MTP for speculative decoding (up to 20% acceptance length increase).
  • Context and output: 1,000,000-token context window with 131,072 max output tokens — Z.ai documentation describes this as "solid" and "lossless" for long-horizon work.
  • Thinking effort: Two levels exposed at the API: Max (default) and High (explicit reasoning_effort="high"). Thinking can be disabled with enable_thinking=false.
  • Agent compatibility: Day-one support for Claude Code, Cline, OpenCode, Roo Code, OpenClaw, Kilo Code, Crush, and Goose via OpenAI-compatible and Anthropic Messages APIs. In Claude Code, set model to glm-5.2[1m] and auto-compact window to 1,000,000.
  • Benchmarks (post-launch): Terminal-Bench 2.1: 81.0 (vs. GLM-5.1 62.0, Claude Opus 4.8 85.0). SWE-bench Pro: 62.1 (vs. GLM-5.1 58.4). Arena Code Arena Frontend: #2 open model, matching Opus 4.8 no-thinking with GLM-5.2 Max mode.
  • Licensing and availability: MIT-licensed open weights on Hugging Face and ModelScope (BF16 and FP8). Together AI serverless endpoint: zai-org/GLM-5.2. Z.ai Coding Plan: all tiers, live June 13. Standalone API and chatbot: June 16.
  • Deployment: SGLang (v0.5.13+), vLLM (v0.23.0+), Transformers (v4.51.2+), KTransformers, Unsloth. Ascend NPU support via vLLM-Ascend, xLLM, SGLang.
  • Fine-tuning: Slime (v0.3.0+), ms-swift (v4.4.0+) for SFT, PPO, GRPO.

Analysis

In our view, GLM-5.2 crosses a usability threshold for open-weight agentic coding. The combination of 1M context, configurable thinking, and drop-in agent support means teams can evaluate it against their actual repositories today — not against synthetic benchmarks. The Terminal-Bench and SWE-bench gains over GLM-5.1 are substantial (19 and 3.7 points respectively), and the Arena ranking suggests the model holds its own in realistic coding workflows. The MIT license removes the friction of Apache-2.0 attribution for commercial embeddings. For production adoption, the key variables are inference cost at Max thinking effort and long-context KV cache behavior under sustained multi-hour sessions — both measurable in a staging environment.

The distribution-first launch through the Coding Plan is a deliberate signal: Z.ai wants developers stress-testing the model in real harnesses before the benchmark narrative hardens. That aligns with how practitioners actually adopt tools — by running a refactor or migration task end-to-end. The two thinking-effort levels are a practical lever: High for latency-sensitive edits, Max for planning-heavy refactors. We recommend starting with Max (default) and only dropping to High if throughput becomes a bottleneck.

Unknown

  • GLM-5.5 existence and specs: No official announcement, parameter count, architecture changes, or training details exist. Any claims about GLM-5.5 are rumors or extrapolation.
  • GLM-5.5 release timeline: No dates, quarters, or roadmap commitments from Z.ai.
  • API pricing changes: Current Together AI serverless pricing applies; Z.ai standalone API pricing (launched June 16) has not been published in a stable, referenceable form. Coding Plan quotas per tier (Lite/Pro/Max/Team) are documented but subject to change.
  • Long-horizon reliability at scale: While 1M context is "solid" per Z.ai and validated in Terminal-Bench, sustained multi-hour agent loops with hundreds of tool calls have not been independently stress-tested in public reports.
  • Ascend NPU vs. GPU parity: Deployment frameworks support Ascend, but performance parity (latency, throughput, thinking-effort behavior) between Ascend 910B and H100/A100 is not benchmarked in the wild.
  • Fine-tuning recipes for agentic workflows: Slime and ms-swift support exists; optimal hyperparameters for coding-agent fine-tunes are not published.

Our take

GLM-5.2 is the first open-weight model that feels production-ready in a general coding agent harness — not a benchmark special. Treat the Coding Plan as your evaluation environment: run a real repo-scale refactor this week. If it holds, the MIT weights give you a self-hosting path that closed models cannot match. GLM-5.5 is a distraction until Z.ai speaks.

Series: 1. GLM-5.2: What Actually Shipped and Why It Matters · 2. The 2026 Roadmap: Zhipu's Release Cadence and Strategy · 3. GLM-5.5 Rumors vs. Official Silence: Parsing the August 2026 Reports · 4. Developer's Guide: What to Treat as Confirmed vs. Unknown in the GLM Ecosystem · 5. GLM-5.3 Ships: Post-Training Coding and Emergent Cyber Defense · Zhipu GLM

Sources