The apps that route the most traffic to GLM-5.2 on OpenRouter are Hermes Agent, Claude Code, pi, Kilo Code, and OpenClaw — a lineup made almost entirely of open-source coding and agent tools. That is worth more than a benchmark score, because every one of these apps lets its user pick from hundreds of models and switch mid-task. The tokens they push to GLM-5.2 are a revealed preference, not a locked-in default.
Benchmarks tell you what a model can do in a lab. Usage tells you what people trust it with when their own work is on the line. OpenRouter publishes the second kind of data, and for GLM-5.2 it points in one clear direction.
The data
| App | What it is | Tokens on GLM-5.2 |
|---|---|---|
| Hermes Agent | Open-source self-improving agent (Nous Research) | ~1.15T |
| Claude Code | Anthropic's terminal coding agent | ~862B |
| pi | Open-source CLI coding agent | ~244B |
| Kilo Code | Open-source coding agent for VS Code, JetBrains, CLI | ~133B |
| OpenClaw | Open-source personal-assistant agent across messaging apps | ~118B |
Snapshot from OpenRouter's public app ranking for GLM-5.2, mid-2026. Token totals are prompt plus completion summed over the ranking window; OpenRouter merges app aliases and excludes private apps. These figures move week to week, so read the ordering and the order of magnitude as the signal, not the exact number. The live list sits on GLM-5.2's OpenRouter page.
Why this list is unusually honest
Most "top apps" lists are marketing. This one is close to the opposite, and it comes down to where the traffic starts.
Every app on it is an aggregator or an agent that hands the model choice to the user. Kilo Code offers more than 500 models at zero markup. pi carries around 300. Hermes Agent and OpenClaw route through OpenRouter's full catalog of 300-plus. When a developer inside one of these tools sends work to GLM-5.2, nothing is nudging them there — no default, no bundling, no discount that only applies to one model. They picked it against the entire field, and kept picking it.
That is what makes the data useful. It is not GLM-5.2's own app reporting GLM-5.2 usage. It is neutral tools, full of alternatives, showing where the work actually landed.
Who is running it, and why
Hermes Agent is the one to understand first, because it sits at the top here and at the top of OpenRouter's overall agent ranking. It is an open-source agent from Nous Research built to run for a long time: persistent memory across sessions, skills it writes for itself, and a messaging gateway that reaches Telegram, Discord, Slack and around twenty other platforms. Long-horizon agents live or die on two things — holding context and calling tools correctly, hundreds of steps deep. GLM-5.2's million-token window and its tool-calling reliability are a direct fit, which is why so much of Hermes' volume lands there.
Claude Code on this list is the telling one. It is Anthropic's own terminal coding agent, and it speaks to any OpenAI-compatible endpoint, so teams point it at GLM-5.2 when they want the same diff-style editing at a fraction of the token cost. Anthropic's tool, running a Chinese open-weight model, near the top of the chart — that is the revealed-preference argument in a single row.
pi is an open-source CLI coding agent that has processed more than two trillion tokens across some three hundred models. It ranks its own models by real use, and GLM-5.2 earns a top slot for the coding work its users actually do.
Kilo Code is an open-source agent for VS Code, JetBrains and the CLI, with 500-plus models at provider cost and no markup. Its users reach for GLM-5.2 on multi-file feature work and refactors, the tasks where a million-token context earns its keep.
OpenClaw is the outlier in kind — a personal-assistant agent that connects to your messaging apps and takes real actions, from shell commands to browsing to files. Automation like that runs cheap models for the routine beats and a stronger one for the reasoning, and GLM-5.2 is showing up as that stronger one.
What the pattern says about GLM-5.2
Four of the five are coding or engineering agents, and the fifth is a general agent doing tool-heavy automation. None of them are chat toys. That tells you what GLM-5.2 is trusted with in production: software work, and long-running agents that call tools.
It lines up with the model's own numbers — 62.1 on SWE-bench Pro, a million-token context, an MIT license — but the usage data says it more convincingly than any benchmark can. For the full breakdown of where those numbers come from, the GLM-5.2 deep dive has it.
Running GLM-5.2 yourself
Every app above reaches GLM-5.2 the same way: an OpenAI-compatible endpoint and a model id. You can do the same from your own code, or from any of these agents.
Turiloop gives you one key for GLM-5.2 — along with DeepSeek, Kimi and MiniMax — billed pay-as-you-go, with an international card and no Chinese phone number. Point your agent's base URL at api.turiloop.com/v1, set the model to glm-5.2, and you are running the same model these apps run. One key covers every model on the list, and the pricing breakdown shows what a coding workload actually costs.
FAQ
What apps use GLM-5.2? By token volume on OpenRouter, the top apps are Hermes Agent, Claude Code, pi, Kilo Code and OpenClaw — mostly open-source coding and agent tools. Because these apps let users pick any model, their GLM-5.2 usage reflects a real choice among hundreds of alternatives.
Why do coding agents choose GLM-5.2? It pairs strong coding results (62.1 on SWE-bench Pro) with a million-token context and reliable tool calling, which is what agents need to work across many files and many steps. See the best-for-coding guide.
Is OpenRouter's app ranking reliable? It is based on measured token usage (prompt plus completion) per app, with private apps excluded and aliases merged. The absolute numbers shift over time, so it is best read as a directional signal rather than a fixed leaderboard.
How do I run GLM-5.2 in my own agent? Set your agent's base URL to an OpenAI-compatible endpoint like api.turiloop.com/v1, add your key, and set the model to glm-5.2. Most agents on this list — aider, Kilo Code, Roo Code, Claude Code — take exactly those three settings.