GLM-5.2 Review — Zhipu's Open-Weight Coding Flagship
On June 13, 2026, Zhipu AI (Z.ai) released GLM-5.2, its open-weight coding flagship, and followed it on June 17 with a benchmark card that did something the Chinese open tier rarely does on day one: it put up cross-vendor coding numbers and beat GPT-5.5 on every one of them. The weights ship under an unrestricted MIT license, the metered API undercuts the closed leaders by roughly six-to-one, and the headline is a genuine, third-party benchmark rather than a proprietary suite. This review covers what GLM-5.2 is, which of its numbers are measured versus estimated, and where it actually fits.
TL;DR verdict
| GLM-5.2 | |
|---|---|
| Type | Coding-first general LLM (agentic SWE focus) |
| Architecture | Sparse MoE, ~40B active / ~753B total |
| Context window | 1M tokens · up to 131,072 output |
| License | MIT (unrestricted open weights) |
| Pricing | $1.40 in / $4.40 out per 1M · $0.26 cached input |
| Headline number | SWE-bench Pro 62.1 (vs GPT-5.5 58.6) |
| Availability | Hugging Face weights, Z.ai API, GLM Coding Plan, OpenRouter |
| Best for | Cost-sensitive agentic coding, self-hosted SWE agents |
| Caveat | Published coding/agentic numbers only — broad-reasoning cells are estimates |
If you skip the rest: GLM-5.2 is the strongest open-weights coding model released this month and one of the best-value coders, period. It published real SWE-bench Pro, FrontierSWE, and MCP-Atlas numbers that edge GPT-5.5 at a fraction of the cost, and it did it under a permissive MIT license. The honest asterisk is that Zhipu led with coding and agentic benchmarks, not the full standard suite — so its general-knowledge ranking is still interpolation.
What it is
GLM-5.2 is a sparse Mixture-of-Experts model with roughly 753 billion total parameters and ~40 billion active per token. That active-parameter count is the lever that keeps inference cheap enough to justify the pricing, and it is the same architectural pattern the whole frontier open tier — DeepSeek V4 Pro, Kimi, Qwen — has converged on. The jump from the GLM-5.1 predecessor is squarely a coding-and-agentic refresh rather than a from-scratch base model.
The context window is 1 million tokens with up to 131,072 tokens of output, a real step up from the 200K window the earlier GLM-5 line carried, and enough to hold a substantial repository slice plus the running history of a long agentic session. The license is the part that matters most: unrestricted MIT, weights downloadable from Hugging Face. Unlike a subscription-gated closed coder, you can self-host GLM-5.2, fine-tune it, and run it air-gapped — which for a lot of enterprise teams is the entire decision.
What the launch numbers say
Here is where GLM-5.2 separates itself from a typical open-weights drop. Zhipu published a benchmark card on June 17, 2026 built around long-horizon, agentic coding evaluations — the kind that actually correlate with day-to-day developer work — and reported results above GPT-5.5 on each:
| Benchmark | GLM-5.2 | GPT-5.5 | GLM-5.1 | What it measures |
|---|---|---|---|---|
| SWE-bench Pro | 62.1 | 58.6 | 58.4 | Hard, contamination-resistant SWE tasks |
| FrontierSWE | 74.4 | 72.6 | — | Long-horizon software engineering |
| MCP-Atlas | 77.0 | 75.3 | — | Agentic tool use over MCP |
Those are vendor-published, so the usual caution applies — wait for independent confirmation before treating a 3.5-point SWE-bench Pro lead as settled. But two things make this card more credible than most launch decks. First, the benchmarks are the hard, agentic ones (SWE-bench Pro, not the easier Verified or Lite), where headroom is real and gaming is harder. Second, the predecessor anchor is consistent: GLM-5.1 scored 58.4 on SWE-bench Pro, so a move to 62.1 is a believable single-generation gain rather than an implausible leap.
What is not in the card: MMLU-Pro, GPQA Diamond, Aider Polyglot, tau-bench, the full standard public suite. That is why, in our models leaderboard, GLM-5.2’s swe_bench cell carries the published SWE-bench Pro 62.1 (matching how the GLM-5.1 row was recorded), while its general-reasoning cells are conservative estimates anchored to GLM-5.1 — MMLU-Pro and GPQA nudged up modestly, coding and agentic axes (HumanEval, Aider, tau-bench) lifted in line with a coding-first refresh, and the multimodal cell left empty because this is a text/code release.
What it costs
On Zhipu’s first-party API:
- Input: ~$1.40 per 1M tokens
- Cached input: ~$0.26 per 1M tokens
- Output: ~$4.40 per 1M tokens
There is also a GLM Coding Plan flat subscription starting around $18/month for the entry tier, aimed at developers who want a predictable bill rather than metered tokens. The framing that traveled furthest at launch — open-weights GLM-5.2 matching or beating GPT-5.5 on long-horizon coding “for one-sixth the cost” — comes from this gap: GPT-5.5 lists at $5/$30 per 1M, so GLM-5.2’s $1.40/$4.40 is a roughly 6× swing on blended cost for a model that, on the published coding benchmarks, lands ahead. And because the weights are MIT-licensed, the metered API is a floor, not a ceiling on savings — run your own GPUs and the per-token cost disappears entirely. The calculator on the leaderboard lets you plug in your own token mix against the closed leaders.
How it compares
GLM-5.2’s real competition is the open coding tier and the value-priced edge of the closed tier, not the absolute frontier. Against Kimi K2.7-Code, released the day before on June 12, GLM-5.2 is the broader generalist of the two and, unlike Kimi, shipped with real cross-vendor coding numbers instead of only proprietary ones — a meaningful credibility edge this week. Against DeepSeek V4 Pro (SWE-bench Verified 80.6) it competes on a different benchmark family, but both occupy the same “frontier-adjacent open weights, priced to undercut” niche. Against closed leaders like Claude Opus 4.8 and the Mythos-class Claude Fable 5, GLM-5.2 is far cheaper and self-hostable, but those models still hold the lead on the hardest reasoning and broad-knowledge benchmarks — the gap the open tier has been closing on coding specifically while the closed leaders keep the reasoning crown.
The pattern worth noticing: the open-weights coding tier is now shipping on a near-weekly cadence — GLM-5.2 and Kimi K2.7-Code landed one day apart — each drop leapfrogging the last on coding benchmarks. If your workload is “edit real code across a large repo, cheaply, possibly on your own hardware,” that competition is working entirely in your favor.
Who should care
- Teams running self-hosted coding agents: This is the headline use case. MIT weights, a 1M context window, and benchmarks tuned to long-horizon agentic work. Pull it from Hugging Face and A/B it against your current open coder.
- Cost-sensitive agentic pipelines: If you are paying closed-model API rates on a high-volume coding workload, GLM-5.2’s roughly 6× cost advantage plus the self-host option is worth a serious pilot — see multi-agent pipelines for where a cheap, strong coder slots into a larger workflow.
- Anyone benchmarking open vs closed: GLM-5.2 is a genuinely useful data point because it published hard coding benchmarks. Still, confirm the SWE-bench Pro lead with third-party runs before you rank it above GPT-5.5 in production planning.
- Teams that need vision or broadest reasoning: Look elsewhere. GLM-5.2 is a text/code model; for multimodal work or the hardest reasoning, a frontier closed model is the better fit — the LLM Benchmark Comparison 2026 covers how to weigh that trade.
FAQ
What is GLM-5.2? Zhipu AI’s open-weight coding flagship, released June 13, 2026. A sparse MoE (~40B active, ~753B total) with a 1M-token context window, up to 131,072 tokens of output, and an unrestricted MIT open-weights license.
How much does it cost? About $1.40 per 1M input tokens, $4.40 per 1M output, and $0.26 cached, on the Z.ai API. A flat GLM Coding Plan starts around $18/month, and MIT weights mean self-hosting is also an option.
Is it better than GPT-5.5 at coding? On the benchmarks Zhipu published it edges GPT-5.5 — SWE-bench Pro 62.1 vs 58.6, FrontierSWE 74.4 vs 72.6, MCP-Atlas 77.0 vs 75.3 — at roughly one-sixth the cost. Those are vendor numbers; GPT-5.5 still leads on broad reasoning.
Does it have published SWE-bench scores? Yes — SWE-bench Pro 62.1, published June 17, 2026. It did not publish the full standard suite (MMLU-Pro, GPQA, Aider, tau-bench), so those remain estimates anchored to GLM-5.1.
Can I self-host it? Yes — the weights are on Hugging Face under an unrestricted MIT license, so you can run, fine-tune, and deploy it on your own infrastructure.
Continue reading
- AI Models Leaderboard — GLM-5.2 versus 60+ models on benchmarks, pricing, and context window, with a cost calculator.
- Kimi K2.7-Code Review — the open coding model released one day before GLM-5.2, and how their launches differ on benchmark transparency.
- GPT-5.6 Review (Sol, Terra & Luna) — OpenAI’s late-June preview, and the cheaper tier GLM-5.2’s pricing is pushing against.
- Claude Code vs Cursor vs Codex — where an open coding model like this one actually plugs into a coding workflow.
- LLM Benchmark Comparison 2026 — how to read coding benchmarks and self-reported numbers without getting fooled.
- All Reviews — index of every head-to-head review on the site.