GLM 5.2 supports Amp, dcode, and Next.js workflows after Composio tops 41 tool tasks

• composio's 41-task eval put GLM 5.2 at 40 of 41 agentic tool tasks, ahead of Claude Opus 4.8 and GPT-5.5 at 39 each.
• Toolmakers moved fast: LangChain's dcode demo showed GLM 5.2 as a selectable model with a pasted API key flow, while sqs on Amp plugin modes said web support for the GLM plugin was being feature-flagged.
• In app-layer coding work, rauchg on Next.js called GLM "excellent" at Next.js, and rauchg's follow-up pointed developers to trusted inference origins for glm-5.2-fast served with Wafer.
• Access is already getting productized: WesRoth on ZCode said Z.ai shipped an official desktop environment for GLM 5.2, and cline on ClinePass added discounted access to GLM plus other open-weight coding models inside CLI and IDE workflows.

You can check Braintrust's GLM-5.2 eval, browse Hyperbrowser's open-source browser harness, and even watch the ClinePass setup clip. The weirdly fast part is how much of the rollout happened through tooling, not a single flagship app: dcode setup in one tweet LangChain's dcode demo, Amp plugin flags in another sqs on Amp plugin modes, and an official ZCode desktop client right behind them WesRoth on ZCode.

Composio's benchmark used 41 real-tool tasks across GitHub, Jira, and LaunchDarkly. Its most specific example was not a codegen trick, it was a state-checking mistake: GLM 5.2 inspected LaunchDarkly's pending approvals list, while the two closed models stopped at the visible on/off flag state composio's 41-task eval.

That makes the win look more like agent discipline than raw benchmark theater. composio on deeper checks framed the differentiator as knowing when to stop and inspect more state.

The dcode path was about as short as these things get: download the app, pick GLM 5.2, paste an API key, start building LangChain's dcode demo.

Amp looked a little more experimental, but more interesting. According to sqs on Amp plugin modes, web support for plugin modes was about to ship behind a feature flag, with amp plugins update or amp plugins add --auto-update @amp/glm-52-mode as the refresh path.

The same Amp thread exposed a product wrinkle: subagents could still reach models a user was not subscribed to, and sqs said fixing that cleanly might require compromising the product behavior sqs on Amp plugin modes.

Guillermo Rauch, Vercel CEO, wrote that GLM is "excellent" at Next.js, citing both evals and his own hands-on use rauchg on Next.js. In a follow-up, rauchg's follow-up pointed to trusted inference origins for glm-5.2-fast served in partnership with Wafer.

That lines up with the broader cost-performance pitch around GLM's coding use. wafer_ai's cost chart priced 100 million tokens at $192.50 for GLM 5.2, versus $1,000 for Opus 4.8 and $1,125 for GPT-5.5 in Wafer's benchmark mix.

Z.ai's official answer to all this routing chaos is ZCode, a desktop coding environment for macOS, Windows, and Linux with BYOK support, existing GLM subscription support, and a 1.5x quota bump for GLM Coding Plan subscribers WesRoth on ZCode.

A separate thread from hyperbrowser's browser post showed the next obvious extension: GLM 5.2 inside a cloud browser loop. Hyperbrowser said its harness reads page text, asks the model for one next action, executes that action in a real browser, and repeats, with the code published publicly at agent-web-index.