Perplexity Adds Z.ai's GLM 5.3 to Power Deep Research Inside Computer

Perplexity's autonomous agent product just picked up a new model in its lineup. GLM 5.3 is now selectable inside Perplexity Computer, where it will handle long-context and multimodal agent workloads. In Perplexity's internal testing, the newer Z.ai model outscored GLM 5.2 on WANDR, the benchmark the company built specifically to measure evidence-backed research agents.

For anyone who has not used Computer, this is the product Perplexity shipped earlier this year as its bet on agent execution rather than better answers. It is an AI agent system that coordinates across dozens of underlying models to interpret screens, reason about tasks, and execute multi-step workflows. Computer breaks a request into discrete steps, assigns each to the appropriate model, and interacts with applications through web or desktop interfaces or APIs. Vision models read the screen, language models handle task logic, and memory systems track state across sessions.

Why swap in GLM 5.3

Computer's whole design premise is that different sub-tasks want different models, so adding a stronger option for one slot changes what the whole system can do. GLM 5.3 targets the deep research slot: long documents, many sources, evidence chains that need to hold up.

That is what WANDR measures. It is an internal Perplexity benchmark for wide-and-deep research inside Computer. Agents have to search broadly enough to find every qualifying entity and dig deep enough to back every claim with evidence. The benchmark contains 500 research tasks that require 170,495 source-backed records across three difficulty tiers. Unlike single-answer evals, WANDR re-fetches every cited page and checks each claim against the underlying evidence, a much harsher test of grounding than the usual multiple-choice format.

What GLM 5.3 brings to the party

GLM 5.3 is Z.ai's newest flagship, and it fits the profile Computer needs. Key specs worth knowing:

• Released mid-August 2026, built on the same base weights as GLM 5.2 with every improvement coming from post-training. It has a 1M token context window and up to 128K output tokens.
• Terminal-Bench 3.0 jumps from 4.6 to 28.3 and DeepSWE v1.1 from 46.2 to 66.9 versus GLM 5.2, so long-horizon coding and tool-use tasks are where the gains show up.
• A separate GLM-5.3-Flash variant is a mixture-of-experts model with 320B total parameters and 18B active per token, a 1,048,576-token context window, and image and video input, released under an MIT license with weights on Hugging Face.
• The training stack includes IndexShare for long-context processing, SAO for reinforcement learning on long-horizon tasks, and slime, Z.ai's open-source post-training framework. System work raised end-to-end RL throughput on long-horizon coding tasks by more than 2.3x.

GLM 5.3 was tuned for the workload Computer throws at its research sub-agent: read a lot, plan across many steps, and use tools without losing the thread. That translated into a better WANDR score than GLM 5.2 when dropped into the same harness.

Where it sits on the leaderboard

GLM 5.3 is not the top scorer on WANDR overall. In Perplexity's published WANDR results, Perplexity Search as Code leads at 0.363 soft F1 and 0.133 hard F1, Anthropic sits second at 0.249 and 0.072, and every other system tops out at 0.121 soft F1 and 0.035 hard F1. Grok 4.5 previously posted the highest WANDR score inside Computer, landing at the same price range as custom GLM 5.2 with Advisor. The news here is that Z.ai's post-training gains carried over into Perplexity's harness, giving Computer users a stronger cost-efficient option for the research slot.

Who this is for

Perplexity Computer is available exclusively to Perplexity Max subscribers at $200 per month and runs entirely in the cloud, so no powerful local machine is required. Tasks execute in an isolated environment with a real filesystem, browser access, and connections to over 400 applications including Slack, Gmail, GitHub, and Notion. With GLM 5.3 in the model menu, the practical use cases it opens up include:

1. Million-token research jobs where every claim needs to tie back to a source, such as due diligence, literature reviews, and competitive intelligence.
2. Multimodal workflows that mix screenshots, PDFs, and dashboards without a separate OCR pipeline.
3. Long-horizon agent runs where the sub-agent has to plan, retry, and use tools over many steps without dropping context.

If your current Computer setup runs lean on either evidence quality or context length, this is the model to try first.