GLM-5 is the release that opened the current generation of the GLM model family. It arrived with materially stronger coding fundamentals than the 4.x line, more reliable multi-step execution, and noticeably better behaviour on complex agent tasks where a plan has to survive contact with real tool output. Everything the 5 series is now known for — long-horizon reasoning, agentic engineering, open weights under MIT — starts here.
Because the model authors publish their full comparison table for GLM-5.2 and GLM-5.1 only, GLM-5 has no benchmark column of its own. This page is honest about that: it gives you the spec sheet and positioning that were published, plus the nearest published reference column, rather than inventing numbers. If you need hard scores to make a decision, the flagship page is the right place to look.
No sign-up, no account, no credits. Type a prompt and get a streamed answer.
Response will stream here...
This playground runs a free open-source LLM via OpenRouter. It is not GLM-5.2 and does not use GLM weights — it's here so you can test a live model while you read the published GLM benchmark data below.
The step from the 4.x line to GLM-5 was about reliability rather than headline capability. Multi-step execution is the part of agentic work that quietly decides whether a run finishes: a model that reasons beautifully but drops a step, forgets a constraint or re-does work it already completed will not survive a fifty-turn trajectory. GLM-5 tightened that behaviour, and it did so alongside better coding and better handling of complex agent tasks, which is why the two releases that followed could focus on horizon length rather than basic dependability. Read the 5 series as three moves: GLM-5 made long trajectories reliable, GLM-5.1 made them long, and GLM-5.2 made them very long by rebuilding the attention stack around a shared indexer.
The published comparison table covers GLM-5.2 and GLM-5.1. The column below is GLM-5.1's — the closest published reference point to the GLM-5 generation. No benchmark column has been published for GLM-5 itself, and none is invented here.
Treat the column above as an upper-bound sanity check rather than a score for GLM-5. It shows where the generation immediately after GLM-5 landed: 95.3 on AIME 2026, 86.2 on GPQA-Diamond, 71.8 on MCP-Atlas and 58.4 on SWE-bench Pro, with long-horizon rows still weak. If you are choosing between GLM models on the basis of measured performance, the only two options with published columns are GLM-5.2 and GLM-5.1, and GLM-5.2 wins every row of that head-to-head.
Comparison on Terminal-Bench 2.1 (Terminus-2), as published by the model authors.
The chart tracks the 5-series trajectory against the open-weight field on Terminal-Bench 2.1: GLM-5.1 at 63.5 and MiniMax M3 at 65.0 sit in the same band, Qwen3.7-Max reaches 75.0, and GLM-5.2 breaks out at 81.0. GLM-5 predates all of these measurements, so its place on that curve is inferred from positioning rather than measured — another reason to move to the flagship page if you need a number you can defend in a design review.
GLM-5 follows GLM-4.7 and precedes GLM-5-Turbo, GLM-5.1 and GLM-5.2. Its direct descendant in spirit is GLM-5-Turbo, which took the same base and tuned it for dynamic, long-chain agent scenarios. For production work today the family points you at GLM-5.2; GLM-5 matters mostly as the point where the current generation's design decisions were set.
The 1M-token context flagship. It is the open-source state of the art on coding and long-horizon evaluations, with explicit Non-Thinking, High and Max effort levels.
A 5-series base model tuned for dynamic, long-chain agent scenarios where the plan changes repeatedly and the trajectory has to stay executable.
Read the GLM model benchmarks, then put a real model to work. The playground above is free and needs nothing from you; if you want a fuller AI toolkit, our partner's free tier starts here.