In late March 2026, Zhipu (Z.ai) announced through its official release notes that GLM-5.1 is rolling out to Coding Plan subscribers. The most notable part of the update is not the decimal in the version number but a single figure: support for autonomous coding runs of up to 8 hours.
Eight hours is a dividing line. The typical coding agent cycle has been measured in minutes — read the request, change code, run tests, report back. When a task can run for a full working day, the engineer’s role shifts from issuing step-by-step instructions to delegating in the morning and reviewing in the evening.
What an 8-Hour Run Actually Tests
Long-horizon autonomy stresses a model on several fronts: context has to stay coherent across a long chain of tool calls, errors have to be detected and recovered by the agent itself, and a failed test midway cannot be allowed to reset the whole run. Writing 8 hours into a product spec is a claim that these problems are past the usability threshold, not just research demos.
The engineering side is tested just as hard. A continuous run amplifies token consumption, tool side effects, and file changes, so the cost of any small mistake compounds over time. That changes acceptance criteria as much as it changes model capability.
From Open-Source Flagship to Subscription Perk
Shipping GLM-5.1 to Coding Plan subscribers first extends the path Zhipu set with GLM-5 in February: model capability leans toward long-horizon agents, while the business model leans toward developer subscriptions. Coding is the agent scenario with the clearest willingness to pay, and turning flagship capability into a subscription perk is the shortest route from model progress to cash flow.
Notes for Developer Teams
- Add “can it resume after an interruption” to acceptance criteria, not just single-run benchmarks
- Longer runs mean larger error costs; code review and test gates have to rise accordingly
- A Coding Plan shifts from cheaper tokens toward an autonomy quota, and cost models need recalculating
Turning an 8-hour autonomous run from a demo into a subscription spec is one step. Whether it holds up under real workloads is a question the coming weeks of subscriber feedback will answer.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.