Z AI launches AutoClaw for local OpenClaw runs with no API key

• Z AI launched AutoClaw, a way to run OpenClaw locally with “no API key required,” support for “any model you like,” and a default path to GLM-5-Turbo; the company says the setup is fully local so “your data never leaves your machine” in the launch post.
• The companion model, GLM-5-Turbo, is a proprietary text-only model with a 200K context window and first-party API access; according to Artificial Analysis, it is “deeply optimized for the OpenClaw scenario.”
• On evals, Artificial Analysis says GLM-5-Turbo trails open-weights GLM-5 (Reasoning) on its overall Intelligence Index, 47 vs. 50, but the GDPval result puts it ahead on GDPval-AA agentic task work, 1503 vs. 1408.
• That tradeoff is not purely cost-driven: the token-efficiency result says GLM-5-Turbo used fewer output tokens than GLM-5 on the full index run, but the benchmark thread still estimates a slightly higher effective eval cost because API pricing is higher.

AutoClaw is the concrete product change here: Z AI says OpenClaw can now be downloaded and started locally, without provisioning an API key first, and with model choice left open rather than locked to a hosted default in the launch post. The launch copy is unusually explicit about the operating model — “fully local” and “your data never leaves your machine” — which makes this less a model release than a deployment path for agent workflows that need local execution.

The same post positions GLM-5-Turbo as the default fit for that workflow. Z AI describes it as optimized for tool calling and “multi-step tasks” in the launch post, and the demo video AutoClaw demo shows a local terminal start-up followed by a successful tool-calling interaction. The practical implication is simple: AutoClaw lowers the setup friction for running OpenClaw on a laptop or workstation, while still letting teams swap in another compatible model if they do not want Z AI’s hosted option.

Artificial Analysis' benchmark thread shows a narrower profile than a generic “best model” claim. GLM-5-Turbo scores 47 on its Intelligence Index, behind GLM-5 (Reasoning) at 50, with weaker results on TerminalBench, CritPt, and HLE. The same thread says it does slightly better on GPQA and IFBench, but the bigger distinction is elsewhere: on GDPval-AA, which tracks agentic real-world work tasks, GLM-5-Turbo posts a 1503 Elo versus 1408 for GLM-5 in the GDPval result. That lines up with Z AI’s positioning around OpenClaw rather than broad reasoning leadership.

There are real tradeoffs. Artificial Analysis says GLM-5-Turbo scores worse on AA-Omniscience, at -15.1 versus +2.0 for GLM-5, which it interprets as weaker knowledge reliability and more hallucination risk in the benchmark thread. It is also not the cheaper version of the same model: while the efficiency note says it used about 94M output tokens on the full eval versus roughly 109M-110M for GLM-5, the thread says higher token prices still push effective run cost slightly above GLM-5. For engineers, that makes GLM-5-Turbo look less like a straight upgrade and more like an OpenClaw-tuned option with a specific agentic sweet spot.