Z.ai releases GLM-5.2 AI model that rivals Claude Opus and beats GPT-5.5 on coding benchmarks

Z.ai pitches GLM-5.2 for long-running software engineering tasks

The open-source model combines a one million-token context window with architectural updates aimed at lowering the cost of repository-scale AI coding. Z.ai has released GLM-5.2, an MIT-licensed open-source AI model designed for long-running software engineering tasks, as the Chinese company seeks to challenge proprietary coding models on cost and performance. The company said GLM-5.2 ranked just behind Anthropic's Claude Opus 4.8 on FrontierSWE, a long-horizon coding benchmark, trailing it by 1%. Z.ai said the model also edged out OpenAI's GPT-5.5 by 1%. Z.ai said GLM-5.2 supports a one million-token context window with up to 131,072 output tokens, positioning it for agentic coding workflows that require reasoning across large codebases.

Z.ai's open-weights GLM-5.2 beats GPT-5.5 on multiple long-horizon coding benchmarks for 1/6th the cost

Today, Chinese AI startup Z.ai (formerly Zhipu AI) announced the immediate release of GLM-5.2, a 753-billion parameter open-weights large language model (LLM) engineered specifically to dominate "long-horizon" autonomous coding and engineering tasks. Available immediately on Hugging Face, the Z.ai API, and more than 20 third-party coding environments, the model boasts a highly stable 1-million-token context window alongside enterprise subscription tiers starting at just $12.60 per month. In excellent news for cost and security-conscious businesses, z.ai has released GLM-5.2's core weights under an unrestricted MIT open-source license, allowing enterprises to download the model freely from Hugging Face, customize or fine-tune it to their liking, and run it potentially locally or via virtual machines for only the cost of their compute and electricity. This is an increasingly appealing option for enterprises, as state-of-the-art American proprietary models face an uncertain and potentially interrupted regulatory future, following the Trump Administration's export control directive last week prohibiting foreign nationals from using Anthropic's new Claude Fable 5 model (which that company responded to by taking the models in question entirely offline for all users). For enterprise technical decision-makers, z.ai's GLM-5.2 provides a highly capable path to host frontier-level AI locally, entirely bypassing the geographic fencing and commercial limitations. IndexShare re-uses one indexer for every four sparse attention layers, reducing compute needs Under the hood, GLM-5.2 operates with 753 billion parameters and introduces a major architectural optimization called "IndexShare". In standard massive language models, recalculating attention mechanisms across long documents is computationally exorbitant. IndexShare solves this by reusing the identical indexer across every four sparse attention layers. At the maximum 1-million-token context length, this single innovation reduces per-token compute FLOPs by a massive 2.9 times. The model also features an upgraded Multi-Token Prediction (MTP) layer for speculative decoding, which boosts accepted token length by up to 20% during inference. Additionally, Z.ai has implemented flexible, selectable "Thinking Modes". Users can toggle the model's reasoning effort between "Max," designed to push the limits of logical problem-solving, or "High," which strikes a careful balance between high-end performance and latency-sensitive token efficiency. State-of-the-art benchmarks for an open model, and matching, even beating proprietary leaders on some categories On industry-standard third-party benchmark tests, GLM-5.2 performs above most open source flagship models, even DeepSeek v4 and scores near or above its closed-weights rivals, OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.8. The model particularly shines in agentic tool use and long-horizon software engineering tasks: * SWE-bench Pro: GLM-5.2 scored 62.1, decisively beating GPT-5.5 (58.6) and its own predecessor, GLM-5.1 (58.4). * FrontierSWE (Dominance): Designed to test long-horizon task completion, GLM-5.2 hit 74.4%, surpassing GPT-5.5 (72.6%) and finishing in a near-tie with Claude Opus 4.8 (75.1%). * MCP-Atlas: On this tool-usage evaluation, GLM-5.2 achieved a 77.0, outscoring GPT-5.5 (75.3) and performing just shy of Claude Opus 4.8 (77.8). * Humanity's Last Exam (w/ Tools): When equipped with external tools, GLM-5.2 reached a score of 54.7, coming out ahead of GPT-5.5 (52.2) and tracking closely behind Claude Opus 4.8 (57.9). * PostTrainBench & SWE-Marathon: In extended, multi-hour engineering workloads, GLM-5.2 consistently topped GPT-5.5, scoring 34.3% against GPT-5.5's 25.0% on PostTrainBench, and 13.0% against GPT-5.5's 12.0% on SWE-Marathon. While GLM-5.2 trails Claude Opus 4.8 and GPT-5.5 slightly on raw Terminal-Bench 2.1 scores (81.0 versus 85.0 and 84.0, respectively), it significantly outscores Google's Gemini 3.1 Pro (74.0). Beyond traditional coding metrics, GLM-5.2 took an impressive first place on the crowdsourced design task benchmark Design Arena, beating out even the aforementioned state-of-the-art Claude Fable 5 with an ELO score of 1360. Furthermore, the impact of Z.ai's new selectable "thinking modes" is clearly visible in the data: under the "Max" effort level, GLM-5.2 pushes to peak intelligence, but utilizes nearly 85k output tokens per task. Switching to the "High" effort setting sacrifices only a few points in performance while effectively halving the required token output, providing a crucial optimization lever for latency-sensitive applications. Available via Coding Plans and API To operationalize the model, Z.ai launched the GLM Coding Plan, aiming squarely at developer workflows rather than simple chat interfaces. The plan offers out-of-the-box support for third-party U.S. and global agentic coding harnesses and tools including Claude Code, OpenClaw, Cline, Kilo Code, Crush, and Factory, among others. The Coding Plan pricing tiers (when billed annually) are highly competitive: * Lite: $12.60 per month ($151.20 per year starting in the 2nd year), geared toward lightweight iteration on small repositories. * Pro: $50.40 per month for day-to-day development on mid-sized repositories, offering 5x the usage allowance of the Lite plan. * Max: $112.00 per month for heavy workloads, offering 20x the Lite usage and dedicated resources during peak hours. For enterprise developers integrating the raw model into their own applications, Z.ai's API pricing undercuts its Western rivals significantly while matching the exact rates of the previous GLM-5.1 generation. GLM-5.2 API access is priced at $1.40 per million input tokens and $4.40 per million output tokens, making it a mid-priced model globally, but about Sorted by total cost (input + output) from least to most expensive. Pricing shown is standard pay-as-you-go pricing per 1 million tokens. To further optimize costs for long-context workloads, Z.ai offers a cached input rate of just $0.26 per million tokens, alongside a limited-time offer for free cached input storage. The stark contrast between open-weights innovators and proprietary Western labs has not gone unnoticed by the developer community. On X, prolific AI observer Lisan al Gaib (@scaling01) argued that "frontier labs are absolutely scamming you on API pricing". The post noted that while massive open models like the 744-billion-parameter GLM-5.2 charge $4.40 per million output tokens and DeepSeek-V4-Pro (1.6 trillion parameters) charges just $0.87, proprietary models demand heavy premiums: Anthropic's Sonnet 4.6 and Opus 4.8 charge $15.00 and $25.00 respectively, while OpenAI's GPT-5.5 costs $30.00 for output. Highlighting that open-model developers are operating profitably without relying on the newest "fancy Blackwell chips," the commentator suggested that leading proprietary labs are "probably at 90%+ margins at this point". The beauty of the unmodified MIT License for enterprise use The most disruptive aspect of the GLM-5.2 release is its licensing. Z.ai released the model's weights under an MIT open-source license, establishing it as a "Pure Open" system. The company's technical documentation explicitly notes that this license guarantees "no regional limits" and allows "technical access without borders". For enterprise technology leaders, an MIT license means the software can be used, modified, and commercialized without paying royalties or adhering to restrictive "acceptable use" governance policies common to dual-use licenses. It allows engineering teams to host frontier-level AI on their own sovereign infrastructure, entirely eliminating vendor lock-in. Warm reception among AI developers and toolmakers The developer reaction to the release has been immediate and overwhelmingly positive. The team behind Kilo Code confirmed day-one integration, posting on X: "GLM-5.2 runs in Kilo Code on day one. The 1M context window and Max effort mode are both live. Point your config at it and go!". Open-source coding environment Cline IDE echoed this sentiment on X, noting the economic advantage: "GLM-5.2 is the first open-weights model to cross 80% on Terminal-Bench, and beats every other open model available. It also beats Gemini, making it a frontier-level model for a fraction of the cost. Open weights is back. This model is a game changer. Available in Cline now!". Similarly, rival open source coding desktop agent Eigent AI also tested the model's new capabilities on complex agentic workflows, noting on X: "threw a real long-horizon task: research 30 companies across 6 sectors of the AI infrastructure stack, structure it into JSON, then build an interactive HTML report... where 5.2 pulls ahead: -> plans...".

China's Z.AI Releases GLM-5.2: A Model That Rivals Claude Opus -- Using Zero Nvidia Chips

Unsloth AI already released 2-bit GGUF quantizations that shrink the model from 1.51TB to 238GB. You'll still need 256GB of RAM or VRAM -- but at that point, you can run it. Z.ai dropped GLM-5.2 on June 16, promising top level performances, beating its already advanced GLM 5.1. The Beijing-based lab, which has been on the U.S. Entity List since January 2025, appears to be benefiting from growing concerns over America's approach to AI. Over the past week, the ban on Anthropic Fable and the release of this new model have helped drive zAI's stock up 90%, sending it to a new all-time high. GLM 5.2 has the numbers to back up the hype. On FrontierSWE -- a benchmark that evaluates whether an AI agent can complete open-ended technical projects measured in hours, covering systems optimization, large-scale code construction, and applied ML research, scored by dominance rate -- GLM-5.2 hit 74.4 against Claude Opus 4.8's 75.1. It edged out GPT-5.5 at 72.6. On SWE-bench Pro, which tests autonomous resolution of real-world GitHub issues scored as a pass rate, GLM-5.2 scored 62.1 to GPT-5.5's 58.6 -- and cleared its predecessor GLM-5.1's 58.4 by a wide margin. The quality jump makes it the best open-source model to date in the Artificial Analysis Intelligence Index, which aggregates the results of 9 different scores to assess the general quality of an AI model. OpenRouter's benchmarks put it in the same category as the now banned Claude Fable 5. The hardware used to achieve this feat is another interesting part of the story. GLM-5.2 was trained on Huawei Ascend chips -- no Nvidia anywhere in the pipeline. Emad Mostaque, founder of Stability AI, estimated total training costs at around $25 million, 80% of that in post-training, which would make it extremely cheap when compared against its peers. As Decrypt reported earlier this year, Z.ai was already training image models on Huawei's Ascend Atlas servers without a single American chip. GLM-5.2 takes that infrastructure further -- a 744-billion-parameter mixture-of-experts model with a genuine 1 million-token context window, five times the 200K limit on GLM-5.1, and an MIT license that means no government directive can flip the access switch. Tokens are the chunks of tet a model can read and generate whereas Parameters are the number of internal settings and values that determine how a model processes information and generates responses Who it's for and what it costs For developers, the context window is the operational shift. Whole-repo navigation, multi-file refactors, and long agentic pipelines that previously required chunking become single-call workflows. API pricing runs $1.40 per million input tokens and $4.40 per million output -- against Claude Opus 4.8's $5 input and $25 output. The Coding Plan starts at around $18 a month and works directly inside Claude Code, Cline, Kilo Code, and most popular agentic environments. Local deployment is also technically possible. Unsloth AI pushed 2-bit GGUF quantizations that compress the model from 1.51TB down to 238GB while retaining ~82% accuracy. Don't get too excited, though. That still means it demands 256GB of unified memory or a matching RAM/VRAM combo -- a maxed M4 Ultra Mac Studio or a workstation with a mid-range GPU and 256GB of system RAM with mixture-of-experts offloading. It's still a lot of money, but at least something that you can buy and run on your house if you really want to. We ran a quick test, asking GLM-5.2 to build our standard game mixing typing mechanics with a shooter. The UI wasn't the prettiest -- other models generated more polished-looking interfaces, but the experience was the most varied: different scenarios across waves, enemy types that shifted, bosses appearing later in the run. It generated more diverse game states than anything else we tested for the same task in a zero shot setup. If you want to play it, it's live in our Itch.io profile. That variance points toward where GLM-5.2 makes the most economic sense. For multi-shot generation workflows and agentic pipelines where output diversity matters more than polish, the math at open-source pricing levels is hard to argue with. For the hardest sustained tasks -- SWE-Marathon, where it scores 13.0 against Opus 4.8's 26.0 -- the gap to the closed frontier is still real, and 13 points wide. Open-source weights are live on HuggingFace under the MIT license. The quantized weights are also available on HuggingFace. GLM Coding Plan subscribers can switch now with the model string GLM-5.2, and it's also available for free testing on z.AI with some usage constraints.