Z.ai’s GLM‑5.1 is an MIT-licensed agent model built for long, tool-heavy workflows—up to eight hours per task—backed by strong benchmarks and paid developer plans.
Key takeaways:
- ai released GLM‑5.1 under an MIT open-source license for commercial use and modification.
- The model is designed for long-duration autonomous work, reportedly up to eight hours on one task.
- GLM‑5.1 is a 754B-parameter Mixture-of-Experts model with a 202,752-token context window.
- Reported tests show large gains in iterative engineering workloads and competitive benchmark scores.
- ai is using a hybrid strategy: open-source GLM‑5.1, while faster Turbo variants remain proprietary.
What Z.ai released and why it matters
Z.ai has released GLM‑5.1 under an MIT open-source license, letting enterprises download, modify, and deploy it commercially. That matters because many advanced agent tools remain closed, which limits customization and self-hosting.
For enterprise teams, MIT licensing also reduces procurement friction. You can run internal security reviews, control data flows, and decide where the model sits in your stack (on‑prem, private cloud, or hybrid) instead of routing everything through a third-party chat product.
The bigger story is the shift from quick replies to long, autonomous execution. Z.ai says GLM‑5.1 can stay on a single task for up to eight hours, turning the model into something closer to a persistent “software worker” than a chatbot.
How GLM‑5.1 is designed for long tool-heavy workflows
GLM‑5.1 is described as a 754‑billion‑parameter Mixture‑of‑Experts model built to maintain goal alignment across workflows with thousands of tool calls. Z.ai claims it can complete up to 1,700 steps in a task, far more than earlier agent systems.
A large context window is central to this: 202,752 tokens, which helps the model keep instructions, logs, code outputs, and constraints in view during long runs.
Z.ai also describes a “staircase” optimization pattern: make incremental improvements, then shift strategy when progress stalls. That is a practical way to avoid the common agent failure mode where the model loops on small changes without rethinking the plan.
Performance highlights and what they imply
Z.ai’s strongest performance claim is about endurance during real engineering loops. In VectorDBBench, GLM‑5.1 reportedly ran 655 iterations and more than 6,000 tool calls to optimize a vector database, improving throughput from 3,547 queries per second to 21,500 after identifying structural bottlenecks.
On KernelBench Level 3, it reportedly achieved a 3.6× geometric mean speedup across 50 machine-learning tasks and maintained progress over extended runs.
Benchmark scores cited include 58.4 on SWE‑Bench Pro, 95.3 on AIME 2026, and 86.2 on GPQA‑Diamond. Even if teams treat benchmarks cautiously, the theme is consistent: GLM‑5.1 is being pitched as an agent that can keep improving a system over hours, not just produce a one-shot answer.
Pricing and how Z.ai is positioning the product
Z.ai is positioning GLM‑5.1 as a developer tool, not a consumer chatbot. It is integrated into a subscription Coding Plan: Lite at $27 per quarter, Pro at $81 per quarter, and Max at $216 per quarter.
For API usage, pricing is listed at $1.40 per million input tokens and $4.40 per million output tokens, with discounts for cached inputs. A separate model, GLM‑5 Turbo, remains proprietary and is optimized for speed.
For teams budgeting agent workflows, token pricing matters because long tasks can generate large tool logs and repeated context. A model that runs for hours needs cost controls like caching, partial context reuse, and clear stopping rules.
Open-source strategy and what to watch next
GLM‑5.1 follows a hybrid strategy: open-source core release, while performance-optimized models remain closed. Z.ai, listed on the Hong Kong Stock Exchange with a stated market value of $52.83 billion, appears to be using open source to expand global developer adoption.
Early feedback cited in the report points to better reliability and long-task execution, including claims of compressing week-long workflows into two days and building complex apps within an eight-hour run. Z.ai also notes remaining challenges such as maintaining coherence and improving self-evaluation when tasks lack clear metrics.
If integrations with popular developer tools (Claude Code, OpenCode, Kilo Code, and others) work smoothly, GLM‑5.1 could become a serious option for teams that want self-hostable agents that run for hours and execute tool-heavy projects end-to-end. The clearest proof will be repeatable results on messy, real production codebases.
