Perplexity Adds GLM 5.3 to Power Computer Agent

Perplexity has expanded the capabilities of its autonomous agent system, Perplexity Computer, by adding Z.ai's GLM 5.3 model to its roster. Available exclusively to Max-tier subscribers for $200 per month, the cloud-based Computer platform coordinates multiple models to execute multi-step workflows across more than 400 application connectors like Slack, Gmail, GitHub, and Notion. The addition of GLM 5.3 targets the platform's deep research slot, offering a highly capable option for tasks that require analyzing massive documents and maintaining long-horizon tool use.

In Perplexity's internal testing, GLM 5.3 outperformed its predecessor, GLM 5.2, on WANDR, a specialized benchmark designed to evaluate evidence-backed research agents across 500 tasks and 170,495 source-backed records. While Perplexity Search as Code leads the WANDR benchmark with a 0.363 soft F1 and 0.133 hard F1 score, and Anthropic follows with 0.249 soft and 0.072 hard F1, GLM 5.3 provides a highly cost-effective alternative. Grok 4.5 previously held the highest WANDR score inside Computer, but the new Z.ai model offers a strong budget-friendly option. The model's improvements are driven entirely by post-training on the same base weights as GLM 5.2, utilizing Z.ai's open-source post-training framework, slime, alongside IndexShare for long-context processing and SAO for reinforcement learning. This stack boosted reinforcement learning throughput on long-horizon coding tasks by more than 2.3 times.

Released in mid-August 2026, GLM 5.3 features a 1-million-token context window and supports up to 128,000 output tokens. Its specialized training yielded massive benchmark gains, with Terminal-Bench 3.0 scores jumping from 4.6 to 28.3 and DeepSWE v1.1 scores rising from 46.2 to 66.9 compared to GLM 5.2. Additionally, a separate GLM-5.3-Flash mixture-of-experts variant has been released under an MIT license on Hugging Face, featuring 320 billion total parameters, 18 billion active parameters per token, a 1,048,576-token context window, and multimodal support for image and video inputs. For practitioners, these upgrades translate to a more reliable agent for complex, multi-step workflows, such as conducting literature reviews or running long-horizon coding tasks without losing track of the underlying evidence.