GLM-4.5 Guide: The Release That Set Z.ai's Agentic Direction

If you want to understand why the current GLM models look the way they do, start with GLM-4.5. It shipped in July 2025, and the decisions it made have carried through every Z.ai release since: a large mixture-of-experts model, a small active slice, one model that can both think and answer quickly, and a benchmark story built around agents rather than chat.

In late 2026 it is no longer the model to reach for by default. But it is still listed on Z.ai's API and on OpenRouter, and still a clean reference point for judging what later releases actually changed.

What GLM-4.5 introduced

Z.ai's technical report describes GLM-4.5 as a mixture-of-experts model with 355B total parameters and 32B active per token, trained on 23 trillion tokens across several stages and then refined with expert iteration and reinforcement learning. It was released alongside a smaller sibling, GLM-4.5 Air, at 106B total and 12B active.

The weights are MIT-licensed, with FP8 and base-model releases alongside the BF16 checkpoint. Z.ai's documentation lists a 128K context window and up to 96K output tokens, text only. Its API also supports function calling, context caching and JSON output.

The headline idea was unification. Earlier open models tended to split into a reasoning variant and a fast chat variant. GLM-4.5 put both behaviours in one set of weights, switchable per request.

Hybrid thinking in practice

GLM-4.5 has a thinking mode for multi-step reasoning and tool use, and a non-thinking mode for immediate replies. On Z.ai's API this is the thinking.type parameter, set to enabled or disabled. When you self-host with vLLM or SGLang, thinking is on by default and you pass enable_thinking: False in the chat template arguments to turn it off.

The practical consequence is that one deployment can serve two kinds of traffic. A classification call or a quick rewrite goes through with thinking off and costs a fraction of the output tokens. A debugging task with tool calls goes through with thinking on. Reasoning models explained covers when the extra tokens actually help; the short version is that they help on multi-step problems and are wasted on lookups.

If your client does not expose the switch, you are paying for thinking on every request. Check that first when a GLM-4.5 bill looks high.

The numbers, and what they do and do not show

The technical report gives three figures that defined the launch:

• TAU-Bench: 70.1%
• AIME 24: 91.0%
• SWE-bench Verified: 64.2%

Across Z.ai's own suite of 12 benchmarks, GLM-4.5 averaged 63.2, which the report places third overall among the models it evaluated, and second on the agentic subset. Z.ai also ran 52 coding tasks inside Claude Code against Claude Sonnet 4, Kimi K2 and Qwen3-Coder. It reported the highest average tool-calling success rate of the four, 90.6%, but in head-to-head task comparisons GLM-4.5 still lost to Sonnet 4 more often than it won: 40.4% wins against 50.0% losses.

Read those numbers as a July 2025 snapshot measured on vendor scaffolds. They told you GLM-4.5 was competitive with the frontier of its moment. They do not tell you how it compares with anything released since, and the vendor ranking was against a field that has moved on considerably. Benchmark contamination explained is worth reading before leaning on any of them.

Serving cost and hardware

Z.ai's pricing page lists GLM-4.5 at $0.60 per million input tokens, $0.11 for cached input, and $2.20 per million output. OpenRouter listed the same headline rates at the time of writing.

Self-hosting is a real undertaking. Z.ai's launch README asked for sixteen H100s or eight H200s for BF16 inference, or eight H100s or four H200s for FP8, with more needed for the full 128K context. That is cluster territory. If one node is your ceiling, Air is the GLM-4.5-generation model that fits.

Should you still run 4.5 rather than 4.6?

Mostly no. GLM-4.6 followed at the end of September 2025 with a 200K context window, a 128K output cap, and what Z.ai describes as more than 30% better token efficiency than GLM-4.5. Z.ai lists both at the same per-token price, so the newer model is the same rate with more room and, by the vendor's account, fewer tokens per task.

There are three honest reasons to stay on 4.5:

• Pinned behaviour. You have an evaluated prompt set and a regression suite tuned to 4.5, and re-validating costs more than the upgrade gains.
• Fine-tunes. You have adapters trained on 4.5 weights. An adapter is tied to the weights it was trained on, so plan to retrain rather than reuse.
• Hardware. You are already serving it and a migration means re-provisioning.

None of those are reasons to start a new project on it.

Migrating off it cleanly

Because GLM-4.6 uses the same inference setup, migration is mostly a model-id change. Do it as an experiment rather than a swap:

1. Freeze twenty representative tasks with their current 4.5 outputs.
2. Run them on 4.6 with the same thinking settings and sampling.
3. Compare completion rate and total output tokens. The efficiency claim is the one that should show up in your logs.
4. Keep 4.5 as a fallback route for a week. Model routing and fallbacks covers the setup.

If the newer model completes the same tasks in fewer tokens, the decision makes itself.