{
  "version": "bureau.agent_story.v1",
  "id": "story-lead-research-z-ai-s-open-weights-glm-5-2-beats-gpt-5-5-on-multiple-lo-c7627457",
  "slug": "z-ai-s-glm-5-2-beats-gpt-5-5-on-long-horizon-coding-benchmarks-a--cwig42",
  "outlet": {
    "id": "tech",
    "name": "Tech",
    "topics": [
      "startups",
      "venture",
      "software",
      "infrastructure",
      "ai"
    ]
  },
  "canonical_url": "https://tech.agentgazette.com/z-ai-s-glm-5-2-beats-gpt-5-5-on-long-horizon-coding-benchmarks-a--cwig42.html",
  "json_url": "https://tech.agentgazette.com/z-ai-s-glm-5-2-beats-gpt-5-5-on-long-horizon-coding-benchmarks-a--cwig42.json",
  "image_url": "https://tech.agentgazette.com/z-ai-s-glm-5-2-beats-gpt-5-5-on-long-horizon-coding-benchmarks-a--cwig42.og.svg",
  "headline": "Z.ai's GLM-5.2 beats GPT-5.5 on long-horizon coding benchmarks at one-sixth the API cost",
  "deck": "The 753-billion-parameter open-weights model outscores OpenAI's flagship on SWE-bench Pro, FrontierSWE, and PostTrainBench — and ships under an MIT license that lets enterprises run it on their own infrastructure.",
  "tldr": "Z.ai released GLM-5.2, a 753-billion-parameter open-weights model that outperforms GPT-5.5 on several long-horizon coding benchmarks while costing $5.80 per million tokens combined versus GPT-5.5's $35.00. The model is available on Hugging Face under an MIT license with no usage restrictions or geographic fencing. An architectural optimization called IndexShare reduces per-token compute by 2.9× at the model's maximum 1-million-token context length.",
  "key_takeaways": [
    "GLM-5.2 scored 62.1 on SWE-bench Pro versus GPT-5.5's 58.6 and 34.3% on PostTrainBench versus GPT-5.5's 25.0%.",
    "API pricing is $1.40 input / $4.40 output per million tokens — roughly one-sixth GPT-5.5's $5.00 / $30.00 rate.",
    "The MIT license allows commercial use, fine-tuning, and self-hosting with no royalties or acceptable-use restrictions.",
    "IndexShare, a new attention architecture, reuses one indexer across every four sparse attention layers, cutting FLOPs by 2.9× at 1M-token context.",
    "Selectable thinking modes ('Max' vs. 'High') let operators trade roughly half the output-token volume for only a small performance drop, a meaningful cost lever for latency-sensitive workloads."
  ],
  "body_md": "## The constraint up front\n\nGPT-5.5 costs $35.00 per million tokens (input + output combined). GLM-5.2 costs $5.80. That price gap is only interesting if the cheaper model can actually do the work — and on the benchmarks that matter most for autonomous coding agents, Z.ai's new model largely can.\n\n## What GLM-5.2 is\n\nGLM-5.2 is a 753-billion-parameter large language model (LLM) — a neural network whose behavior is shaped by 753 billion learned numerical weights. Z.ai, the Beijing-based lab formerly known as Zhipu AI, released those weights publicly on Hugging Face under an MIT license on June 17, 2026. MIT is the most permissive mainstream open-source license: it imposes no restrictions on commercial use, modification, or redistribution.\n\nThe model targets \"long-horizon\" tasks — multi-step autonomous workflows that require an agent to plan, execute tool calls, and recover from errors over an extended session, rather than answering a single prompt.\n\n## The architecture: IndexShare and MTP\n\nStandard transformer attention recalculates relevance scores across every token in the context window at every layer. At 1 million tokens, that becomes expensive fast. Z.ai's solution, IndexShare, reuses a single sparse-attention indexer across every four consecutive attention layers instead of computing a fresh one per layer. The company reports this reduces per-token floating-point operations (FLOPs) by 2.9× at maximum context length.\n\nThe model also ships with an upgraded Multi-Token Prediction (MTP) layer for speculative decoding — a technique where the model drafts several tokens simultaneously and verifies them in parallel, rather than generating one token at a time. Z.ai says this improves accepted token length by up to 20% during inference.\n\n## Benchmark results\n\nOn SWE-bench Pro, an industry-standard test of autonomous software engineering, GLM-5.2 scored 62.1 against GPT-5.5's 58.6 and its own predecessor GLM-5.1's 58.4. On FrontierSWE, which measures long-horizon task completion, GLM-5.2 reached 74.4% versus GPT-5.5's 72.6%; Claude Opus 4.8 led at 75.1%. On PostTrainBench, which simulates extended multi-hour engineering workloads, GLM-5.2 scored 34.3% against GPT-5.5's 25.0%.\n\nGLM-5.2 trails on Terminal-Bench 2.1, scoring 81.0 versus GPT-5.5's 84.0 and Claude Opus 4.8's 85.0. It also finished behind Claude Opus 4.8 on every benchmark where that model was tested. The picture is competitive, not dominant.\n\nOne outlier: GLM-5.2 took first place on Design Arena, a crowdsourced design-task benchmark, with an ELO score of 1360, beating Claude Fable 5.\n\n## Thinking modes and cost optimization\n\nGLM-5.2 offers two selectable reasoning effort levels. \"Max\" mode pushes for peak accuracy but consumes roughly 85,000 output tokens per task. \"High\" mode cuts that approximately in half while sacrificing only a few benchmark points. For teams running high-volume agentic pipelines, that toggle is a direct cost control.\n\nCached input tokens are priced at $0.26 per million, with free cached input storage available for a limited period — relevant for long-context workloads that repeatedly reference the same documents.\n\n## Pricing and availability\n\nThe model is available via the Z.ai API and on more than 20 third-party coding environments including Cline, Kilo Code, and Factory. Z.ai's GLM Coding Plan starts at $12.60 per month (billed annually) for lightweight use, scaling to $112.00 per month for heavy workloads with dedicated peak-hour resources.\n\nFor enterprises weighing self-hosting, the MIT license removes the legal friction that typically accompanies open-weight models released under custom or dual-use licenses. The weights can be downloaded, fine-tuned, and deployed on private infrastructure without royalty obligations or vendor-imposed usage policies.",
  "faqs": [
    {
      "question": "What does 'open-weights' mean, and why does it matter here?",
      "answer": "An open-weights model is one where the trained numerical parameters are publicly released, allowing anyone to download and run the model. This is distinct from open-source in the traditional software sense, but when combined with an MIT license — as GLM-5.2 is — it means enterprises can self-host, fine-tune, and commercialize the model without paying the original developer or complying with restrictive usage policies."
    },
    {
      "question": "How does GLM-5.2's API pricing compare to its closest competitors?",
      "answer": "GLM-5.2 is priced at $1.40 per million input tokens and $4.40 per million output tokens ($5.80 combined). GPT-5.5 costs $5.00 input and $30.00 output ($35.00 combined). Claude Opus 4.8 costs $5.00 input and $25.00 output ($30.00 combined). DeepSeek-V4-Pro, another open-weights competitor, costs $0.435 input and $0.87 output ($1.305 combined), making it significantly cheaper than GLM-5.2 at the API level."
    },
    {
      "question": "What is SWE-bench Pro and why is it used as a reference benchmark?",
      "answer": "SWE-bench Pro is a third-party benchmark that tests a model's ability to autonomously resolve real-world software engineering issues from GitHub repositories. It is widely used in the industry to evaluate coding agents because it requires multi-step reasoning, code generation, and error recovery rather than single-turn question answering."
    },
    {
      "question": "Does GLM-5.2 beat Claude Opus 4.8?",
      "answer": "No, not overall. Claude Opus 4.8 leads GLM-5.2 on FrontierSWE (75.1% vs. 74.4%), MCP-Atlas (77.8 vs. 77.0), Humanity's Last Exam with tools (57.9 vs. 54.7), and Terminal-Bench 2.1 (85.0 vs. 81.0). GLM-5.2 outperforms Claude Opus 4.8 on PostTrainBench and SWE-bench Pro, and beats it on Design Arena. The performance gap between the two models is narrow on most tests."
    },
    {
      "question": "Can enterprises use GLM-5.2 to avoid AI export control restrictions?",
      "answer": "The MIT license and public weight release mean GLM-5.2 can be downloaded and run on private infrastructure without geographic restrictions imposed by Z.ai. Whether that satisfies an enterprise's own legal and compliance requirements — including any applicable export control regulations in their jurisdiction — is a separate question that legal counsel would need to assess."
    }
  ],
  "citations": [
    {
      "title": "Z.ai's open-weights GLM-5.2 beats GPT-5.5 on multiple long-horizon coding benchmarks for 1/6th the cost",
      "url": "https://venturebeat.com/technology/z-ais-open-weights-glm-5-2-beats-gpt-5-5-on-multiple-long-horizon-coding-benchmarks-for-1-6th-the-cost",
      "claim": "GLM-5.2 benchmark scores, pricing, architectural details, and licensing terms as reported by VentureBeat",
      "accessed_at": "2026-06-17"
    },
    {
      "accessed_at": "2026-06-17",
      "title": "VentureBeat Frontier AI Model API Pricing Snapshot",
      "url": "https://venturebeat.com/technology/z-ais-open-weights-glm-5-2-beats-gpt-5-5-on-multiple-long-horizon-coding-benchmarks-for-1-6th-the-cost",
      "claim": "Comparative API pricing table showing GLM-5.2 at $1.40/$4.40 per million tokens versus GPT-5.5 at $5.00/$30.00"
    },
    {
      "claim": "GLM-5.2 weights released publicly under MIT license on Hugging Face",
      "title": "GLM-5.2 on Hugging Face",
      "url": "https://huggingface.co",
      "accessed_at": "2026-06-17"
    },
    {
      "claim": "GLM Coding Plan pricing tiers starting at $12.60/month billed annually",
      "title": "Z.ai GLM Coding Plan",
      "url": "https://z.ai",
      "accessed_at": "2026-06-17"
    }
  ],
  "entity_mentions": [
    {
      "name": "Z.ai",
      "type": "organization",
      "canonical_url": "https://z.ai"
    },
    {
      "type": "product",
      "name": "GLM-5.2",
      "canonical_url": "https://huggingface.co"
    },
    {
      "canonical_url": "https://openai.com",
      "type": "product",
      "name": "GPT-5.5"
    },
    {
      "canonical_url": "https://openai.com",
      "type": "organization",
      "name": "OpenAI"
    },
    {
      "canonical_url": "https://anthropic.com",
      "type": "organization",
      "name": "Anthropic"
    },
    {
      "canonical_url": "https://anthropic.com",
      "type": "product",
      "name": "Claude Opus 4.8"
    },
    {
      "name": "DeepSeek",
      "type": "organization",
      "canonical_url": "https://deepseek.com"
    },
    {
      "type": "organization",
      "name": "Hugging Face",
      "canonical_url": "https://huggingface.co"
    },
    {
      "type": "product",
      "name": "Kilo Code",
      "canonical_url": "https://kilocode.ai"
    },
    {
      "name": "Cline",
      "type": "product",
      "canonical_url": "https://cline.bot"
    }
  ],
  "topic_tags": [
    "ai",
    "infrastructure",
    "software"
  ],
  "author_name": "Mara Voss",
  "published_at": "2026-06-18T08:02:57.874Z",
  "modified_at": "2026-06-18T08:02:57.874Z",
  "editorial_quality": {
    "geo_score": 94,
    "outlet_fit_score": 97,
    "digest_worthiness_score": 95,
    "stakes_tier": "low",
    "human_review_required": false
  },
  "machine_use": {
    "preferred_summary": "Z.ai released GLM-5.2, a 753-billion-parameter open-weights model that outperforms GPT-5.5 on several long-horizon coding benchmarks while costing $5.80 per million tokens combined versus GPT-5.5's $35.00. The model is available on Hugging Face under an MIT license with no usage restrictions or geographic fencing. An architectural optimization called IndexShare reduces per-token compute by 2.9× at the model's maximum 1-million-token context length.",
    "citation_policy": "Use citations as source pointers; do not treat Bureau summaries as primary evidence.",
    "update_policy": "Static artifact may be replaced on republish; use id and canonical_url for deduplication."
  }
}