GLM-5.2

A third-party guide updated September 9, 2026, separates independently verified GLM-5.2 benchmark results from Zhipu's self-reported numbers across categories. GLM-5.2's clearest independently verified result is its #1 finish on Design Arena's Code Categories leaderboard, a blind human-preference test where it ranks ro The guide notes that specific point scores on SWE-bench Pro (62.1), Terminal-Bench 2.1, and FrontierSWE come from Zhipu's own technical report and model card — directionally corroborated by independent evaluators but not reproduced exactly by a third party using Zhipu's precise methodology. The headline framing of 'nea

MorphLLM's August 21, 2026 analysis confirms GLM-5.2's specifications as the prior generation: a 753B-parameter mixture-of-experts model with 40 active parameters, 1M-token context window, and API pricing of $1.40 input / $4.40 output per 1M tokens. The piece notes that the August 14, 2026 GLM-5.3 successor ships with Independent Artificial Analysis data cited in the article scores GLM-5.3 at 60 on the Intelligence Index, tying Kimi K3 at roughly one-fifth the price. For developers using NovitaAI's GLM-5.2 endpoint, the relevant takeaway is that GLM-5.2 retains its $1.40/$4.40 pricing and 1M context while GLM-5.3 remains Z.ai-exclus

ExplainX's launch coverage (August 14, 2026, with updates through August 22) frames GLM-5.2's June release for contrast: MIT-licensed weights were available on Hugging Face almost immediately at launch, whereas GLM-5.3 ships with a staged safety-review gate for both weights and API access. GLM-5.3 is described as post- The article also cites a serving-layer forensic finding that OpenRouter's free stealth model Ox Alpha runs on Z.ai GLM infrastructure (30/30 tokenizer match, shared error code 1214), which is relevant to NovitaAI's continued role hosting GLM-family models via OpenRouter routing. For GLM-5.2 users, the staged rollout of

Novita’s August 7, 2026 guide places GLM-5.2 on a shortlist of hosted open models for agentic coding, alongside Kimi K2.7 Code and DeepSeek V4 Pro. It characterizes GLM-5.2 as a long-context option worth watching, rather than declaring a universal leaderboard winner. For developers choosing a model through Novita, the guide recommends matching the model to operational requirements such as self-hosting, hosted long-context inference, or reliability across extended tool-use loops in a sandboxed agent runtime. The supplied excerpt does not provide Novita-specific pricing, endpoint beh

GLM 5.2 is available on Novita AI with 1M context, 128K max output, function calling, structured outputs, and serverless API access.

Semgrep published a security research blog on June 22, 2026, titled "We have Mythos at Home: GLM 5.2 beats Claude in our Cyber Benchmarks," evaluating GLM-5.2 on their IDOR benchmark using the same prompt and dataset used for frontier coding agents. According to the report, when models were given nothing but a prompt, This benchmark result is relevant to NovitaAI's hosted zai-org/glm-5.2 variant because it demonstrates that the underlying Z.ai model — the same one Novita serves — is competitive with leading proprietary coding agents on real security workloads, particularly IDOR vulnerability detection. Developers using NovitaAI's GL

Independent commentary by Simon Willison dated June 17, 2026 confirms that Z.ai released GLM-5.2's full open weights under an MIT license on June 16, 2026. The model is described as a 753B-parameter Mixture-of-Experts network with 40 active parameters (1.51TB total), text-input-only, and features a 1 million token cont The article notes GLM-5.2 is also ranked 2nd on the Code Arena WebDev leaderboard behind Claude Fable 5 for front-end web development including agentic coding workflows, and is available via OpenRouter from approximately 9 different providers, almost all charging $1.40/M input and $4.40/M output — a price point directl

Z.ai (formerly Zhipu AI) released GLM-5.2 on June 16, 2026, as a 753-billion-parameter open-weights large language model engineered for long-horizon autonomous coding and engineering tasks, according to VentureBeat reporting. The model ships under an MIT open-source license on Hugging Face, is also available via the Z. Architecturally, GLM-5.2 introduces an optimization called IndexShare, which reuses a single indexer across every four sparse attention layers, reducing per-token compute FLOPs by approximately 2.9x at the maximum 1-million-token context length. It also features an upgraded Multi-Token Prediction layer for speculative

A Hacker News discussion (772 points, ~52 days before the current date) centers on a tweet from Z.ai founder Jie Tang announcing GLM-5.2 as "Fully Open, Frontier Intelligence Belongs to Everyone." According to the announcement, GLM-5.2 is Zhipu's most capable open-source model to date, supports a 1M-token context windo From a NovitaAI routing perspective, GLM-5.2 is delivered through OpenRouter's provider network, where Willison notes GLM-5.2 is hosted across 9 providers, though the excerpt does not explicitly name NovitaAI among them for GLM-5.2 (NovitaAI previously hosted GLM-4.6). Pricing on OpenRouter is reported at roughly $1.40

Techsy.io's open-source LLM leaderboard, last re-verified on July 19, 2026, places Z.ai's GLM-5.2 at the top of its July 2026 ranking and characterizes it as a 744-billion-parameter Mixture-of-Experts model with 40 billion active parameters, released in June 2026 under an MIT license. The article cites 91.2% on GPQA Di The leaderboard also frames the broader 2026 open-weight landscape, noting that DeepSeek V4 Pro still wins on price, Qwen3.6 wins on license freedom, and Gemma 4 12B beats the prior year's 27B flagship at lower memory cost, with Kimi K3 reportedly ahead of Claude Opus 4.8 on Artificial Analysis. The article is a single

NIST's Center for AI Standards and Innovation published an assessment of Z.ai's GLM-5.2 on July 8, 2026, reporting it was probably the most capable open-weight model at its June 16, 2026 release. CAISI found GLM-5.2's overall capabilities are similar to GPT-5.2 (December 2025) and its cyber capabilities are similar to Anthropic's Opus 4.6 (February 2026). The evaluation also flagged mixed safeguards performance: GLM-5.2 allowed assistance with agentic cyber exploit development, blocked fewer sensitive biological questions than reference U.S. models, but appeared potentially more robust against agent hijacking and jailbreaking than other evaluated PRC open-weight models. The report notes that prompt-based safeguards can still be circumvented when the open-weight model is self-hosted, a key caveat for developers.

An aggregator page presents all 17 published GLM-5.2 benchmark results grouped into reasoning, coding, and agentic categories, explicitly disclaiming that the figures are republished as published by the model authors and that the site does not run the evaluations. The long-horizon trio — FrontierSWE (74.4%, up to 20-ho Generation-over-generation deltas versus GLM-5.1 show uneven gains: +3.7 on SWE-bench Pro (62.1 vs 58.4), +17.5 on Terminal-Bench 2.1 (81.0 vs 63.5), +28.2 (2.6×) on DeepSWE (46.2 vs 18.0), +12.8 on ProgramBench (63.7 vs 50.9), +5.0 on MCP-Atlas, +7.5 on Tool-Decathlon (48.2 vs 40.7), and +9.5 on HLE (40.5 vs 31.0). Th