
Ox Alpha Model: How to Switch to GLM-5.3-Flash
By Olivia Hart
MyClaw Editorial
MyClaw
Run Best-in-Class AI Agents Now
Run OpenClaw or Hermes Agent on managed MyClaw hosting, so your AI agent stays online, updated, and ready for real work.
AI Takeaway
- What was Ox Alpha? Ox Alpha, sometimes written “0x Alpha” with a zero, was Z.ai’s anonymous preview of GLM-5.3-Flash on OpenRouter and OpenCode.
- Is Ox Alpha still free? The free preview ended. OpenRouter’s old listing is still visible, but
stealth/ox-alphahas no active endpoints. The released model is available through paid API providers. - Which model ID works now? Use
z-ai/glm-5.3-flashin the OpenRouter API,glm-5.3-flashin Z.ai’s direct API, oropenrouter/z-ai/glm-5.3-flashin OpenClaw with OpenRouter configured. - What should you test first? Try one job that uses the features you need, such as fixing a UI bug from a screenshot and running the relevant test. Check that your provider accepts the image and that your agent can run the test.
Why Ox Alpha Disappeared From Model Pickers
Ox Alpha appeared on OpenRouter on August 20, 2026. OpenRouter called it a stealth model because the developer’s identity was kept private during the preview. Z.ai later confirmed that it had used the same codename on OpenRouter and OpenCode to gather feedback before the August 26 launch of GLM-5.3-Flash.
The historical listing still describes what the preview could do. OpenRouter’s endpoint registry shows zero active endpoints for that ID, so a saved Ox Alpha entry can still appear in your client even though it can no longer serve a request. If your agent stopped working after the preview, check both the default model and any model pinned to the job or conversation. Updating the default may leave those overrides in place.
What GLM-5.3-Flash Can Do

Coding With Screenshots and Tools
Suppose a layout change cuts off the checkout button on a narrow screen. Give the agent the screenshot, viewport dimensions, and affected component. It can inspect the layout, propose a patch, and use a browser or test runner to check the result if you’ve given it access. A concrete acceptance check is that the button remains visible at the failing width and the existing checkout test still passes.
Z.ai documents text, image, video, and file inputs, with text output and function calling. Check which inputs and API features your provider supports, then make sure your client passes them through. A working image request confirms the model can receive the screenshot. A separate tool call confirms the agent can do something with it, such as opening the relevant file or running the test.
How Flash Differs From GLM 5.3 and FlashX
Z.ai lists roughly 1 million tokens of context and up to 128K output tokens for Flash. The large context window can hold the relevant code, screenshots, and logs for one investigation. Check your provider’s effective limit before relying on the full window, especially if your agent adds chat history and tool output to each request.
For a difficult text-only task, such as tracing a bug across several services, the GLM 5.3 Flash vs. GLM 5.3 comparison can help you decide when to test the flagship. If Flash already produces a good result and response time slows down interactive work, try GLM 5.3 FlashX where available. It’s the speed-focused variant and has its own API ID.
Which Model ID to Use Now
Replace the Preview ID in Your Client
| Where you use the model | Model ID | What to do |
|---|---|---|
| Old OpenRouter preview | stealth/ox-alpha | Replace this retired ID in saved settings. |
| OpenRouter API | z-ai/glm-5.3-flash | Send this in the request’s model field. |
| Z.ai direct API | glm-5.3-flash | Use this with your Z.ai endpoint and credentials. |
| OpenClaw with OpenRouter | openrouter/z-ai/glm-5.3-flash | Include the OpenClaw provider prefix. |
In OpenCode, connect your provider with /connect, then choose its GLM-5.3-Flash entry from /models. That selection applies to the current session. Set the model field in opencode.json or opencode.jsonc if you want new work to use the same provider and model by default.
In OpenClaw, once OpenRouter is configured, run openclaw models set openrouter/z-ai/glm-5.3-flash to change the default. Start a fresh session and send a short text request to check the connection. If the request fails, verify the provider credentials and model ID before tracing the rest of the workflow.
If text works but the agent skips a required tool call, check whether the client included the tool definition and whether the chosen provider supports it. For requests you send directly to OpenRouter, set provider.require_parameters to true to restrict routing to providers that support the parameters you send. That gives you a clearer failure when no eligible provider is available.
Check Price, Tool Support, and One Real Run
On November 16, 2026, OpenRouter listed GLM-5.3-Flash at $0.0399 per million input tokens and $0.0784 per million output tokens through Relace. Through Together, the listed rates were $0.15 and $0.50. Check the current rate, input limits, and tool support for the provider you’ll actually use. For a hosted agent, MyClaw pricing covers the cloud instance. Check your inference charges with the model provider when using your own key. Z.ai also publishes MIT-licensed weights if you plan to run the model on your own infrastructure.
Z.ai documents Flash as a reasoning model with thinking enabled. OpenRouter bills reasoning tokens as output, so use the usage record for the full run when estimating cost. A short final answer can still follow several rounds of reasoning and tool calls.
Compare Flash with your current model on one job. Give both runs the same prompt, files, permissions, and time limit. Judge the result using the same acceptance test, and record elapsed time, failed tool calls, retries, and cost. When comparing OpenRouter runs, record the serving provider so you can reproduce the setup. If you need a baseline, the best model for OpenClaw guide explains how to choose one based on the job’s tool and context needs.
Test GLM-5.3-Flash in a MyClaw OpenClaw Agent

Once the new model passes a trial run, put it in the environment that will run the job regularly. With MyClaw’s OpenClaw hosting, your private agent can check a staging site on a schedule even when your computer is off.
Step 1: Select GLM-5.3-Flash
Open your instance’s model settings and choose GLM-5.3-Flash if it’s listed. If you’re bringing your own key, use the Day 2 quick start to set up credentials for the supported provider that offers it. Confirm the active model in a new agent session.
Step 2: Verify the Workflow in the Hosted Agent
Run the screenshot task in the hosted instance using the same repository revision and test as your trial. Check that the agent can reach staging, open the right files, and run the test there. Review the diff and confirm that the checkout button is visible at the failing viewport width.
Step 3: Schedule the Check and Keep a Working Model
Schedule the diagnostic part of the workflow. Give the agent the staging URL, viewport size, and condition to check. Have it report the affected page and attach a screenshot when the button is clipped. Set the schedule’s time zone and trigger one run by hand. Check that the scheduled run can reach staging with its own settings and that the report arrives at the intended destination. Keep a supported fallback where your settings allow it, or record the previous model so you can switch back manually.
Conclusion
If you’re replacing Ox Alpha in several places, migrate one job at a time so you know which change affected the result. Start with a recurring workflow that already has a clear pass condition. Check the first few runs for failures you would otherwise have to catch manually, such as a missed alert or an incomplete test report. Set GLM-5.3-Flash as the default for that job once it meets your quality and cost targets. If it does well on visual checks but struggles with deeper fixes, keep it on the checks and route the fixes to your current model.
Skip the Setup, Run Best-in-Class AI Agents Now
Launch a managed OpenClaw or Hermes Agent workspace in minutes, with always-on hosting, updates, and support handled by MyClaw.