GLM-5.1 Guide: The Eight-Hour Agent Claim, Six Months On

GLM-5.1 arrived on 7 April 2026 with a claim that was different in kind from most launch material. Z.ai did not lead with a quality score. It led with duration: the model, it said, can work independently on a single task for up to eight hours, planning, executing, testing and fixing without a human stepping in.

Six months and two successors later, that claim is the right lens for deciding whether GLM-5.1 still belongs in your stack. Duration is the property it was built for, and duration is also what its 200K context window ultimately limits.

The eight-hour claim and what it measures

An eight-hour run is not a benchmark. It is a description of behaviour: the model keeps making useful progress after the point where earlier models plateau and start repeating themselves. The Hugging Face model card puts it as sustaining optimisation over hundreds of rounds and thousands of tool calls.

Z.ai's documentation gives two worked examples. In one, the model ran 655 autonomous iterations on a vector database and raised query throughput to 6.9 times the starting version. In another, on KernelBench Level 3, it reports a 3.6 times geometric-mean speedup against 1.49 times for torch.compile in max-autotune mode.

Both examples share a shape worth noticing: a measurable objective, a fast feedback signal, and a task where many small improvements accumulate. That is where long-horizon persistence pays. A task with a vague goal and no automated check does not benefit from eight hours; it just spends them.

SWE-Bench Pro at launch, and since

The number that got GLM-5.1 noticed was 58.4 on SWE-Bench Pro, which Z.ai's comparison put ahead of GPT-5.4, Claude Opus 4.6 and Gemini 3.1 Pro at the time. The model card also lists:

• Terminal-Bench 2.0 — 63.5 under Terminus-2, 69.0 best self-reported
• NL2Repo — 42.7
• CyberGym — 68.7
• BrowseComp — 68.0
• HLE — 31.0 without tools, 52.3 with tools
• GPQA-Diamond — 86.2

These are vendor-run figures, and the lead was a launch-day lead. GLM-5.2 followed in June with 62.1 on SWE-Bench Pro, and GLM-5.3 followed in August with its own set of agentic results. Vendor-claimed vs independent benchmarks is worth reading before treating any of these as settled.

200K is the constraint that dated it

Z.ai documents a 200K-token context and a 128K maximum output for GLM-5.1. GLM-5.2 and every GLM-5.3 model ship 1M.

For an eight-hour agent, 200K is tight. A long run accumulates file reads, test output and stack traces, and at some point the harness has to compact history to keep going. Compaction is lossy, and the detail it drops is often the detail the model needed thirty turns later. So the model built for duration runs into a window that forces it to forget partway through.

That does not make long runs impossible — compaction and external memory work, as context compaction strategies covers — but it means part of the eight hours is spent re-reading what it already knew. GLM-5.1 vs GLM-5.2 goes through exactly what the 1M jump changed.

Same price as its successors

This is the awkward part. Z.ai's pricing page lists GLM-5.1, GLM-5.2 and GLM-5.3 at identical rates: $1.40 per million input tokens, $0.26 for cached input and $4.40 per million output. OpenRouter listed GLM-5.1 somewhat lower at the time of writing, at about $0.97 input and $3.04 output per million tokens; rates there depend on which hosts are serving it, so check before you budget.

When an older model costs the same as a newer one with a five-times larger window, the default answer for new work is obvious. GLM-5.1 does not win on price at Z.ai, and it does not win on context anywhere.

Reasons to keep running it anyway

There are still legitimate ones:

• Reproducibility. If you published or shipped results on GLM-5.1, the old model is the only way to reproduce them. Pinning matters more than upgrading for evaluation baselines.
• A tuned harness. Prompts, tool schemas and stop conditions tuned against one model often degrade on its successor in ways that take a week to diagnose. If your loop works, migration is a cost, not a free upgrade.
• Self-hosting. The weights are MIT-licensed. The model card lists 754B parameters and support in SGLang, vLLM, xLLM, Transformers and KTransformers. An existing deployment you trust is a reason to stay.
• Reasoning you can turn off. GLM-5.1 still accepts thinking.type set to disabled. GLM-5.3 does not.

How to test a long-horizon claim on your own work

If you are deciding between GLM-5.1 and a successor, test the property GLM-5.1 was sold on.

1. Pick three tasks with an automated success signal — a benchmark script, a test suite, a performance target.
2. Run each with no turn limit and log the score after every iteration.
3. Plot score against turn number. A model with genuine endurance keeps improving; one without it flattens and starts repeating changes.
4. Note where compaction kicks in. If the curve breaks at that point on GLM-5.1 and not on a 1M model, the window is your bottleneck, not the model.

The general method is in long-running agents. The takeaway is simple: for new work, start on a GLM-5.2 or 5.3 model; keep GLM-5.1 only where a pinned baseline or a tuned harness makes switching more expensive than the gain.