China AI Bench
6 min read

GLM-5.2: the open-source SOTA, warts and all

ReviewBenchmarks Deskby Benchmarks Desk, Testing & rankings desk

Zhipu's GLM-5.2 leads the open-source field on Artificial Analysis at 51 points, tops the Code Arena for available models, and publishes its weak spots: 45-minute reasoning runs and a 5% knowledge lag. We compiled the numbers.

Zhipu released GLM-5.2 on June 17, 2026 with an MIT license, and it is currently the best open-weight model on Artificial Analysis's aggregate index at 51 points, ahead of every other open model and level with the frontier labs' top-three on that chart.

What makes this release worth a compiled piece is that Zhipu published the warts. The official docs list the slow inference, the knowledge gap, and the tool-call edge cases. Rare for a launch, and the most useful kind of documentation for people who actually buy tokens.

What shipped

GLM-5.2 is a 744B-parameter MoE, about 40B active, with a 1M token context window and 128K max output [1]. Text-only, unlike Kimi K3's native vision [1][2]. It landed on Hugging Face and ModelScope under MIT on June 17, four days after Zhipu's June 13 statement responding to tightened Anthropic export restrictions: frontier intelligence shouldn't belong to a few companies [3].

The 1M context is the headline feature, and it's real in a way that earlier long-context claims weren't. Zhipu's demo ran a single instance through an 880K-token session to build a three-platform app (web, mobile, mini-program) [1][4]. Zhipu also showed a 740,000-line server log root-cause analysis and a cross-contract conflict check over four contracts, with 100% tool-call correctness and JSON validity in the demo [5]. Vendor demo, we label it as such.

Where it ranks

MetricGLM-5.2Notes
AA Intelligence Index51, open-source #1top-3 overall alongside Anthropic and OpenAI [3]
FrontierSWE~1% behind Opus 4.8beats GPT-5.5 and Opus 4.7 [1][3]
SWE-Marathon~13% behind Opus 4.8long-horizon agentic coding [1]
Code Arena#1 among available modelsmillion-user blind test [3]
DesignArena1360 ELO, #1first open model at the top of both coding and design [3]
Community engineering runs3 A-grades out of 5 public projectson par with Opus 4.8, with fewer tokens burned [6]

All of these are third-party evals or vendor-reported runs we cross-checked against the community thread. The engineering note is the one I like: on the same failing project, Opus 4.8 burned 564 tool calls and 260K output tokens, GLM-5.2 used 557 calls and 170K tokens [6]. Comparable capability, less waste.

The warts Zhipu published

The official docs and release coverage list three honest weak spots:

  • Reasoning speed. The same long task took GLM-5.2 45 minutes versus Opus 4.8's 33 [5].
  • Knowledge benchmarks. HLE and GPQA trail the frontier by about 5% [5].
  • Multi-step instructions occasionally drop separators, a small but real annoyance in agent pipelines [5].

The speed number matters more than it looks. For agentic workflows, 45 vs 33 minutes on a long task is not a benchmark detail, it's a billable-hours difference. Zhipu published it anyway. That's the tell of a vendor that believes capability sells on its own.

Pricing and the product push

GLM-5.2 API pricing is $1.4 per million input tokens and $4.4 per million output tokens, snapshot July 31 [7]. That's mid-pack: cheaper than Kimi K3 ($3/$15 [8]) and Opus, pricier than DeepSeek's Flash ($0.14/$0.28 [9]).

The consumer angle is the GLM Coding Plan, which starts at $18/month (Lite) and goes up to roughly $96/month (Max) [10]. For context, that positions it as roughly a tenth of Claude Code subscription pricing for a comparable experience, which is how Zhipu's own materials pitch it [10]. It supports 20+ coding tools including Claude Code, Cursor, Cline, and OpenCode [11].

One more data point worth knowing: Cloudflare's Workers AI published optimization work on long-context MoE models including the GLM family, KV cache quantization plus weight compression, claiming 41% higher throughput and 30% lower inference cost [12]. When cloud infrastructure vendors start engineering for your model, that's adoption.

Our take

GLM-5.2 is the open-source leader on the metric that matters most to us, the AA aggregate, and it got there without winning a single headline race. It doesn't top Frontend Code Arena like K3, it doesn't undercut everyone on price like DeepSeek Flash, and it doesn't have K3's 2.8T scale. It just ships a 1M context that works, publishes its own weak spots, and prices like a professional tool rather than a loss leader.

The pricing decision is the interesting part. Zhipu raised prices across the Coding Plan while the market was cutting them (OpenAI's Luna went down 80% the same week [13]). That's a bet that capability and reliability, not cheapness, is the moat. The 45-minute reasoning time is the honest cost of that bet: it's a slower, more deliberate model, and the docs say so.

Risks and what we're watching

  • The knowledge gap. 5% behind on HLE and GPQA is survivable for coding buyers and fatal for knowledge-work buyers. We're tracking whether the next GLM release closes it.
  • Speed. 45 minutes on long tasks will filter out latency-sensitive agent workloads regardless of price.
  • The price trajectory. Coding Plan went up over 30% across three rounds before 5.2, and overseas API pricing rose 67-100% [5]. Zhipu is testing how much the open-source market will pay for SOTA.

Bottom line

GLM-5.2 is the open-source SOTA on the aggregate index, with a genuinely usable 1M context, honest documentation of its slow spots, and mid-pack pricing. It's the model you pick when you need long-context reliability more than raw speed, and the vendor is betting you'll pay for it. The leaderboard row has the numbers.

Same number on our leaderboard → GLM-5.2 row

Sources

  1. Zhipu official docs (GLM-5.2 spec, FrontierSWE/SWE-Marathon), docs.bigmodel.cn/cn/guide/models/text/glm-5.2 (published 2026-06-15, captured 2026-08-06, official)
  2. IT Home (K3 native vision), ithome.com/0/982/259.htm (published 2026-07-27, captured 2026-08-06, media report)
  3. Juejin (AA 51 open-source SOTA, Code Arena #1, DesignArena 1360, June 13/17 timeline), juejin.cn/post/7654122741078671402 (published 2026-06-24, captured 2026-08-06, community)
  4. CSDN (880K-token three-platform demo), blog.csdn.net/weixin_50937681/article/details/162128481 (published 2026-06-20, captured 2026-08-06, community)
  5. iHeima (vendor demo numbers, weak spots: 45min vs 33min, HLE/GPQA -5%, price history), m.163.com/dy/article/KVFSBKR605118I96.html (published 2026-06-15, captured 2026-08-06, media report)
  6. Zhihu (5 public projects, 3 A-grades, token usage comparison), zhihu.com/question/2049161661567723164 (published 2026-06-14, captured 2026-08-06, community)
  7. 36Kr (GLM-5.2 API pricing $1.4/$4.4), 36kr.com/p/3919319636290946 (published 2026-07-31, captured 2026-08-06, official pricing, via media)
  8. Ollama (K3 pricing $3/$15), ollama.com/library/kimi-k3 (published 2026-08-01, captured 2026-08-06, official)
  9. Puter dev blog (V4 Flash $0.14/$0.28), developer.puter.com/tutorials/deepseek-api-pricing/ (published 2026-06-17, captured 2026-08-06, official)
  10. CNBlogs (Coding Plan tiers from $18/mo, 1/10 Claude pricing), cnblogs.com/jiayu1027/p/20507250 (published 2026-06-13, captured 2026-08-06, community)
  11. CodingPlan (plan details, 20+ tools), codingplan.org/plans/zhipu (published 2026-08-01, captured 2026-08-06, community)
  12. OSChina / Cloudflare (KV cache + weight compression: +41% throughput, -30% cost), oschina.net/news/487999 (published 2026-08-05, captured 2026-08-06, media report of vendor engineering post)
  13. 36Kr (Luna -80% same week), 36kr.com/p/3919319636290946 (published 2026-07-31, captured 2026-08-06, media report)

Disclosure: no sponsors, no affiliate links. Compiled from public sources with capture dates; we did not run GLM-5.2 ourselves. Vendor demo numbers are labeled as such.