GLM-5.2 — An Open-Weights Model Built for Long-Horizon Coding
Why This Matters
On June 16, 2026 , Z.ai (formerly Zhipu AI) released GLM-5.2 , a ~753B-parameter open-weights model built specifically for long-horizon agentic coding — the kind of task that runs for hours, spans dozens of tool calls, and accumulates a long, messy trajectory the model has to keep coherent. It ships under an unrestricted MIT license , with a stable 1M-token context window , and it is — per Z.ai's own benchmark table — the highest-ranked open-source model on every long-horizon coding benchmark they report .
The headline that traveled fastest: GLM-5.2 beats GPT-5.5 on multiple long-horizon coding benchmarks at roughly 1/6th the API cost , and lands within a few points of Claude Opus 4.8 — while being something you can download and run on your own hardware. That combination (frontier-adjacent coding capability + MIT weights + no regional restrictions) is the actual story, and it arrives at a moment when proprietary frontier access is getting less predictable, not more (Anthropic pulled Fable 5 / Mythos 5 entirely after a US export-control order the week prior).
This is the same pattern we flagged with deepseek v4 huawei ascend : a Chinese lab shipping open weights that compress the gap to the closed frontier and undercut it on price by an order of magnitude. GLM-5.2 is that pattern aimed squarely at coding agents .
What Got Released
The model
- ~753B total parameters (built on the GLM-5 MoE backbone — 744B total / 40B active), MoE sparse-attention architecture.
- 1M-token context — up 5× from GLM-5.1's 200K. Z.ai's framing is pointed: "A 1M context is easy to claim, but much harder to keep reliable under real engineering pressure." They expanded 1M-context training specifically for coding-agent trajectories (large-scale implementation, automated research, perf optimization, complex debugging).
- Effort-level control — selectable High or Max thinking modes, letting you trade capability against latency/token cost per task. Max pushes peak intelligence at ~85k output tokens/task; High sacrifices a few points while roughly halving token output.
- MIT license, "Pure Open." Weights on HuggingFace and ModelScope. Runs on transformers, vLLM, SGLang, xLLM, ktransformers. No regional limits, no acceptable-use governance, no vendor lock-in.
The architecture — IndexShare
The new trick is IndexShare : every 4 sparse-attention (DSA) layers share one lightweight indexer — the indexer runs on the first of the 4 layers, and its top-k indices are reused for the other 3. This cuts the indexer dot-product and top-k compute in 3/4 of layers, reducing per-token FLOPs by 2.9× at 1M context. They also reworked the MTP layer for speculative decoding (IndexShare + KV-share + rejection sampling + end-to-end TV loss), lifting acceptance length +20% (4.56 → 5.47 on coding scenarios).
The training infra
Post-training ran on slime , Z.ai's agentic-RL framework, using critic-based PPO (single-rollout, token-level advantages) rather than group-relative optimization — a deliberate fit for compaction , where one super-long trajectory gets split into many variable-length sub-traces. They also shipped an anti-hack module : coding RL is uniquely vulnerable to reward hacking (an agent can
curl the reference solution or cat the hidden test cases), so an online rule-filter + LLM-judge guard blocks those calls mid-rollout and returns dummy results instead of aborting the trajectory. The Benchmarks That Matter
| Benchmark | GLM-5.2 | GLM-5.1 | GPT-5.5 | Claude Opus 4.8 | Gemini 3.1 Pro |
| Terminal-Bench 2.1 (Terminus-2) | 81.0 | 63.5 | 84.0 | 85.0 | 74.0 |
| SWE-bench Pro | 62.1 | 58.4 | 58.6 | 69.2 | 54.2 |
| FrontierSWE (dominance) | 74.4 | 30.5 | 72.6 | 75.1 | 39.6 |
| PostTrainBench | 34.3 | 20.1 | 28.4 | 37.2 | 21.6 |
| SWE-Marathon | 13.0 | 1.0 | 12.0 | 26.0 | 4.0 |
| MCP-Atlas (tool use) | 76.8 | 71.8 | 75.3 | 77.8 | 69.2 |
| AIME 2026 | 99.2 | 95.3 | 98.3 | 95.7 | 98.2 |
The shape of it: GLM-5.2 edges out GPT-5.5 on the long-horizon coding benchmarks (FrontierSWE, SWE-bench Pro, PostTrainBench, SWE-Marathon), trails Opus 4.8 by small margins (1% on FrontierSWE, 4 points on Terminal-Bench), and is the first open-weights model to cross 80% on Terminal-Bench . It also took 1st place on the crowdsourced Design Arena (ELO 1360), beating even Claude Fable 5. The improvement over GLM-5.1 is the loudest signal — Terminal-Bench 63.5 → 81.0, FrontierSWE 30.5 → 74.4 — a genuine generational jump in one point release.
Pricing — The Other Half of the Story
| Input / 1M | Output / 1M | Total | |
| GLM-5.2 | $1.40 | $4.40 | $5.80 |
| GPT-5.5 | $5.00 | $30.00 | $35.00 |
| Claude Opus 4.8 | $5.00 | $25.00 | $30.00 |
| DeepSeek-V4-Pro | $0.435 | $0.87 | $1.305 |
- API pricing matches GLM-5.1's exact rates — the capability jump came free.
- Cached input at $0.26 / 1M (with a limited-time free cached-input storage offer) — meaningful for long-context coding where the repo context is re-sent constantly.
- GLM Coding Plan from $12.60/mo (Lite) → $50.40 (Pro, 5× usage) → $112 (Max, 20× usage), with out-of-the-box support for Claude Code, OpenCode, Cline, Kilo Code, Crush, Factory, and Z.ai's own ZCode desktop agent.
- Note the quota mechanics if you subscribe: GLM-5.2 burns quota at 3× peak / 2× off-peak (off-peak billed 1× as a promo through end of September). Peak is 14 –18 Beijing time.
How Significant Is This Compared to Prior Releases?
| Version | Year | Significance |
| GLM-4.5 | 2025 | 355B / 32B active MoE; the "vibe coding" generation |
| GLM-5 | 2026 | Scaled to 744B / 40B active, 28.5T training tokens; "from vibe coding to agentic engineering" |
| GLM-5.1 | 2026 | 200K context; strong open coding model, still trailing the closed frontier |
| GLM-5.2 | Jun 2026 | ~753B, 1M context, IndexShare, effort-level control, MIT — first open model past 80% on Terminal-Bench, beats GPT-5.5 on long-horizon coding |
Where It Fits in Our Workflow
- A real self-hosted coding-agent option. MIT weights + 1M context + 2.9× FLOPs reduction at full context means GLM-5.2 is the strongest open candidate yet for a local agentic-coding tier — work that today routes to Opus or GPT-5.5 but doesn't need to leave our own infrastructure. Pairs with the local-inference thread from DeepSeek deepseek v4 huawei ascend and the Llama 5 distilled-variants question.
- Drop-in for the harnesses we already use. It speaks Claude Code and OpenCode natively (
GLM-5.2[1m]enables the 1M window in Claude Code — note the[1m]suffix convention, same shape as our ownclaude-opus-4-8[1m]). Worth a side-by-side on a real long-horizon task in this tree — a content roll-up or a cross-submodule refactor is exactly the "hours-long, messy trajectory" it's tuned for. - Cost lever for batch coding work. At $5.80 total per 1M tokens (and $0.26 cached input), GLM-5.2 is ~1/6th of GPT-5.5 and ~1/5th of Opus 4.8 for output-heavy agent loops. For cost-tolerant bulk work it's a serious substitute.
- The "effort level" idea is worth stealing conceptually. High vs. Max as an explicit per-task knob — rather than a hidden reasoning budget — is a cleaner mental model than what most frontier models expose. Useful framing when we decide how much compute a given agent task in the tree deserves.