Summary: Zhipu (Z.ai) confirmed in its official release notes in late March 2026 that GLM-5.1 has started reaching Coding Plan subscribers. Let one task stretch across a full working day, and the engineer’s job turns from walking the agent through each step into a morning handoff and an evening review. The update’s real headline is not the extra decimal in the version number but one figure: autonomous coding runs lasting up to 8 hours.
Zhipu (Z.ai) confirmed in its official release notes in late March 2026 that GLM-5.1 has started reaching Coding Plan subscribers. The update’s real headline is not the extra decimal in the version number but one figure: autonomous coding runs lasting up to 8 hours.
That figure splits two eras. A coding agent’s usual loop runs on minutes: take in the request, edit code, execute tests, report the outcome. Let one task stretch across a full working day, and the engineer’s job turns from walking the agent through each step into a morning handoff and an evening review.
What Running 8 Hours Straight Puts to the Test
Long-horizon autonomy strains a model from several directions at once: context must hold together through a long chain of tool calls, the agent has to catch and fix its own mistakes, and one failed test midway must not zero out the entire run. Baking 8 hours into a product spec asserts these problems have crossed the usability threshold — no longer research demos.
Engineering is probed just as hard: an unbroken run magnifies file changes, tool side effects, and token consumption, and the price of a minor slip compounds over time. Acceptance criteria shift as much as model capability does.
An Open-Source Flagship Turned Subscription Perk
Handing GLM-5.1 to Coding Plan subscribers ahead of anyone else follows the route Zhipu charted with GLM-5 in February: capability tilts toward long-horizon agents, business toward developer subscriptions. Among agent scenarios, coding shows the clearest willingness to pay, so recasting flagship capability as a subscription perk is the most direct way to convert model progress into cash flow.
What Developer Teams Should Keep in Mind
- Make “can it resume after an interruption” part of acceptance review instead of leaning on single-run benchmarks
- As runs stretch out, each error costs more, so code review and test gates should move up in step
- A Coding Plan now means an autonomy quota rather than cheaper tokens, and cost models need recalculating
Promoting an 8-hour autonomous run from demo to subscription spec is one step; surviving real workloads is what the next few weeks of subscriber feedback will show.