Cursor adds GLM 5.3 and GLM 5.3 Flash models
Cursor adds native support for Z.ai's GLM 5.3 and GLM 5.3 Flash models, featuring a 1M token context window and on-demand pricing outside standard plans.
The update introduces open weights from Z.ai with agent mode support and usage-based billing outside monthly quotas.
Anysphere has rolled out official support for open-weight GLM 5.3 models in its Cursor code editor, adding both the flagship GLM 5.3 and the lightweight GLM 5.3 Flash. The company announced the release on X on October 5, 2026, alongside updates to its technical documentation. Built by research lab Z.ai (Zhipu AI), both models appear directly inside the editor’s settings menu. Each version supports agent mode, letting the editor read and modify files, run terminal commands, and browse the web.
From manual API setups to the native picker
Until now, developers who wanted to run the GLM family inside Cursor had to configure custom API keys for Z.ai endpoints or route calls through third parties like OpenRouter. That manual process caused setup headaches and steady traffic on Cursor’s community forum. Demand picked up after Z.ai published the weights in late August 2026. By bringing the models straight into the editor’s catalog, Anysphere removed those setup hurdles.
Architecture of GLM 5.3 models
The flagship GLM 5.3 relies on a Mixture-of-Experts (MoE) design, where specialized sub-networks activate dynamically to process each token. The model contains roughly 743 billion total parameters while activating 40 billion per generated token. Its context window spans 1 million tokens (1,048,576 tokens) with three available reasoning tiers, defaulting to High. Z.ai distributed the weights under a permissive license with an added safety audit requirement limited to companies generating over $10 billion in annual revenue.
GLM 5.3 Flash targets lower computational overhead and faster completion times. It holds 320 billion total parameters and activates 18 billion per token. It is also the first model in the GLM-5 family with native vision capabilities, processing images and text under an unrestricted MIT license. To keep memory footprints down during typical coding sessions, Flash relies on hybrid attention mechanisms.
Technical benchmark performance
Evaluation data released by Z.ai and verified in technical write-ups from QDNA and Ofox.ai indicates measurable gains over the previous GLM-5.2 baseline. On the DeepSWE v1.1 software engineering benchmark, the task success rate climbed from 46.2% to 66.9%. On Terminal Bench 3.0, which scores command-line problem solving, the mark rose from 4.6% to 28.3%. In its announcement on X, Cursor highlighted the GLM 5.3 Max variant as the highest-scoring open model on its internal CursorBench 4.0 evaluation suite.
Treat vendor-run numbers with normal caution, as Z.ai ran these tests in its own benchmarking harness. While GLM 5.3 edged out established proprietary models like Claude Opus 4.8 (21.1%) on Terminal Bench 3.0, it remained behind top commercial options such as GPT-5.6 Sol (34.6%).
Pricing and billing details
Within Cursor, both models sit in the metered “Other Models” pool. Queries do not consume fast requests from paid subscription plans; they are billed strictly on demand. Pricing scales linearly across the entire 1-million-token window, with no long-context penalties or cache write charges.
GLM 5.3 costs $1.40 per million input tokens ($0.26 for cached reads) and $4.40 per million output tokens. GLM 5.3 Flash cuts those numbers by roughly 89%: $0.15 per million input tokens ($0.029 on cache reads) and $0.50 per million output tokens. Cursor notes in its documentation that inference runs on partner infrastructure, without naming the hosting providers.