GPT-5.6 Luna can help a local agent inspect a signed-in dashboard when the host client can run the exact gpt-5.6-luna model and call a browser tool. Install Tabbit Browser, sign in to the dashboard yourself, and connect the client through Tabbit CLI. Keep the first run read-only: ask for named changes, evidence URLs, and items that still need checking. GPT-5.6 Luna is distinct from GPT-6 Luna, and neither model alone can see your local browser.
A useful workflow separates three checks: can the provider serve this model to your account, can your client launch the Tabbit skill or CLI, and can the Tabbit Browser profile open the signed-in page? OpenAI lists gpt-5.6-luna for its API and openai.gpt-5.6-luna on Amazon Bedrock. That listing does not prove that a particular client account or local browser connection is ready. OpenAI's current model documentation also lists gpt-6-luna separately. For capabilities and API costs, see the GPT-5.6 family guide.
Key takeaways
Use the exact GPT-5.6 model ID
gpt-5.6-luna; do not select GPT-6 Luna by mistake.Verify model access in the provider and client before configuring browser control.
Skill discovery, local CLI reachability, and website login are separate checks.
Use Tabbit Browser with Tabbit CLI for a local task tab that can reuse the selected Tabbit environment’s existing sign-in.
Ask for a concrete dashboard delta and source URLs. Review before changing or publishing anything.
Check which layer is ready
OpenAI describes GPT-5.6 Luna as a cost-sensitive, high-volume model. Its model page lists text and image input, function calling, and Responses API tools, including computer use. Those are provider capabilities. A third-party client still has to expose the model, run locally, and connect the browser tool. OpenAI model documentation.
| Check | Evidence to look for | What it does not prove |
|---|---|---|
| Provider route | gpt-5.6-luna in OpenAI, or openai.gpt-5.6-luna in Bedrock docs | That your account is enabled or funded |
| Client model picker | Cursor documents a GPT-5.6 Luna entry; select it and confirm a short request in your own account | That the client can reach your local Tabbit Browser |
| Tabbit skill | The client lists the installed tabbit skill or an equivalent skills command | That the local launcher runs |
| CLI connection | A harmless read through the installed Tabbit CLI | That the dashboard is signed in |
| Dashboard read | The task tab shows the expected heading, date range, and values | That the agent’s interpretation is correct |
Do not infer GPT-5.6 Luna availability from GPT-6 Luna, ChatGPT model labels, or a deprecated alias. OpenAI’s API list and your host client’s live picker are the relevant checks. If the model is missing, stop and resolve provider access first. This is also why a client catalog or model directory alone is not evidence of a successful request.
Prepare a local Tabbit task
Open Tabbit Browser and select the profile you use for this dashboard.
Sign in yourself. Open the dashboard, confirm the account and date range, and leave the page visible.
Install the Tabbit CLI Skill in the local agent client. Cursor discovers user skills under
~/.agents/skills/; invoke the skill as/tabbitwhen it appears in the local Agent skill menu. Pi uses/skill:tabbit. Other clients may show/skillsor a picker.Verify the launcher by asking for a read-only page title. Finding the skill and reaching the installed CLI are different things.
Confirm the agent reports the expected page title and current dashboard range before asking it to compare figures.
Tabbit CLI uses a separate task tab, so you can keep working in other browser tabs while the agent reads the dashboard. The parent verification for this batch opened a protected test page before and after a short-lived test login: the second task read the same page in its own task tab. That confirms session reuse in this local test environment, not that every website permits it or that GPT-5.6 Luna performed the test. See how to connect an agent to an existing browser.
A cloud-hosted agent may not reach your computer’s Tabbit Browser. Use it to prepare a comparison checklist, then run the checklist in a local client with a supported Tabbit connection. Do not send session cookies, passwords, or private dashboard contents to a remote service. For a simple way to validate the connection first, use the browser-agent first-task checklist and compare the browser MCP and CLI approaches.
Run a dashboard change check
Choose a dashboard you are authorized to view and a fixed time range. Start with a small, reversible observation task. Ask the agent to compare the visible dashboard against the previous period, list only changes it can point to on the page, and attach source URLs. If a chart does not expose its values or source, ask the agent to mark that item unverified.
Use the Tabbit CLI skill in my current Tabbit Browser profile.
Read the signed-in [dashboard name] for [date range] and compare it with [prior date range].
Do not edit, save, submit, publish, or change filters.
Return a draft named "Weekly dashboard delta — [team] — [date]" with:
- metric and current value
- prior value and direction of change
- the visible label or source URL supporting each item
- anything you could not verify
Stop at the draft. I will review it before sharing or changing the dashboard.Expected output is a short named draft with one line per metric, its visible values, and a source link or page label. The title and values above are placeholders. They are not findings from a live account. Keep sensitive identifiers and customer data out of the prompt unless your organization has approved that use.
| Task result | Human review before proceeding |
|---|---|
| Page title and date range match | Check the profile and selected dashboard |
| Values and labels are visible | Compare each value with the open page |
| The agent infers a cause | Treat it as a hypothesis unless the page shows evidence |
| A filter or date range needs changing | Ask first, then verify the new range |
| The draft is ready to share | Review the full text and links; do not auto-publish |
Leave a named draft for review
The useful output is not an untraceable summary. Keep a named draft in the agent conversation, for example Weekly dashboard delta — Growth — 2026-11-09. Each item should include the metric, both values, the visible label or source URL, and a confidence note only when something could not be checked. You can continue using other Tabbit tabs for source research, then compare the task tab with the draft before sharing it.
OpenAI documents function calling and computer-use tools, and Cursor lists GPT-5.6 Luna with Cursor Agent tools. Those tools do not mean the model can reach your local Tabbit profile by itself. The local Tabbit CLI Skill supplies that connection. Tabbit Browser and Tabbit CLI provide a local browser path; they do not grant dashboard permissions or sign in for you. If the client does not expose GPT-5.6 Luna or cannot invoke the local CLI, resolve that constraint before relying on this workflow.
Common failure points
If the request returns “model not found,” verify the exact provider route and account availability. If the skill appears but no page opens, test the installed CLI and its permissions. If the task reaches a login page, sign in manually in the correct Tabbit profile and retry the read. If a report omits the date range or the visible source, treat it as incomplete.
A community report from July described a gpt-5.6-luna 404 in one Codex setup; OpenAI Support replied with troubleshooting steps in October. That is one client incident, not evidence that the current OpenAI API route is unavailable. The current model page is the source for the API ID and current documented support. For troubleshooting, see Codex and Tabbit CLI setup.
Verdict
Use GPT-5.6 Luna for this browser workflow only after you confirm the exact model in a client that can run locally and reach Tabbit CLI. Start with a read-only comparison, keep a named draft with checkable source links, and make a person approve any change or publication. If you need to compare pricing or model limits, use the GPT-5.6 family guide; this page is about the local signed-in browser workflow.
Frequently asked questions
For a local Codex client, see Codex and Tabbit CLI; the browser session reuse guide and Claude Code setup show other client-specific entry points.
FAQ
Is GPT-5.6 Luna the same model as GPT-6 Luna?
No. These are separate model IDs: `gpt-5.6-luna` and `gpt-6-luna`. Confirm the exact name in the client and provider documentation before running a task.
Can GPT-5.6 Luna open a signed-in website by itself?
No. The model needs a host client and a browser tool. A local client connected through Tabbit CLI can use the active Tabbit Browser environment, where you sign in to the site yourself.
Which provider route is documented for GPT-5.6 Luna?
OpenAI documents the API ID `gpt-5.6-luna` and the Bedrock ID `openai.gpt-5.6-luna`. Account eligibility and client support still need to be checked in your own model picker.
Does the Tabbit skill sign me in?
No. Sign in manually in the target Tabbit Browser profile. Tabbit CLI can then work with that environment’s session when the site permits it.
How do I keep the dashboard task read-only?
Tell the agent not to click save, submit, publish, or change settings. Ask it to return the dashboard name, visible date range, exact changed values, and URLs for the source pages, then compare them yourself.
Can I keep browsing while the task runs?
Yes. Tabbit CLI creates a separate task tab. Keep your other tabs open for research or review, and inspect the task tab before accepting the result.