GPT-5.4 nano can help make a review table of visible support-ticket IDs and links, but it cannot browse a signed-in site on its own. OpenAI documents generic function calling and MCP for the model, while its model page marks built-in computer use as unsupported. An agent client still needs to provide and run a browser tool. A local route to evaluate is Tabbit Browser with Tabbit CLI from a client that supports local tools; in Codex, the installed Tabbit skill is requested with $tabbit. Check the API route, client skill discovery, local CLI reachability, and site login separately. GPT-5.4 nano is also deprecated, with an API shutdown scheduled for April 1, 2027, so keep migration in view.
A Reddit user reports that the model passes their evaluations for structured extraction on clean inputs, but can give confident wrong answers on ambiguous ones. That is one person’s experience, not an accuracy guarantee. It is a good reason to show each ticket ID beside its source URL and make a human check the pair before anyone acts. Reddit discussion
Key takeaways
OpenAI's current API docs use the request ID
gpt-5.4-nano; the CSV/catalog labelopenai/gpt-5.4-nanomay be namespaced for a directory or third-party router.The model supports function calling and MCP, but its model page says built-in computer use is not supported. A client must run Tabbit CLI or another approved browser tool.
Read only the ticket number and its visible URL. Keep names, email addresses, message content, attachments, and account identifiers out of the output.
Use a named, read-only draft. Check every ID-link pair and stop before a ticket action.
OpenAI marks the model deprecated with an April 1, 2027 shutdown date and recommends
gpt-6-luna. Recheck support and rerun the same extraction test before migrating.
Confirm the API model and tool path first
| Route or status | What the current docs say | Decision before a browser task |
|---|---|---|
| OpenAI API | Exact model ID gpt-5.4-nano; function calling, structured outputs, and MCP are listed as supported | Confirm the API account can use the model and the client implements the chosen tool protocol |
| Built-in computer use | GPT-5.4 nano's model page lists computer use as unsupported | Do not expect a native screenshot-and-click tool from this model; provide a separate browser tool through the client |
| Deprecation | Deprecated; API shutdown date April 1, 2027; gpt-6-luna is the recommended replacement | For a new production workflow, check replacement access and repeat the evaluation before the deadline |
| Catalog/provider label | Topic CSV has openai/gpt-5.4-nano | Use the selected provider's documented request ID; do not assume a catalog prefix is accepted by OpenAI's endpoint |
The launch post says nano is API-only. The current model page identifies gpt-5.4-nano, lists function calling and MCP, and separately marks the built-in computer-use tool unsupported. These facts can coexist: a local agent can expose a browser as a custom function or MCP tool, but then the host is doing the browser work. It is not GPT-5.4 nano's native computer-use feature. Read OpenAI's model page, API tools guide, and GPT-5.4 mini and nano announcement before choosing a client.
The API lifecycle matters more than a one-time successful call. OpenAI's deprecations page lists gpt-5.4-nano for shutdown on April 1, 2027 and recommends gpt-6-luna. It also lists the dated nano snapshot in the affected family. Pinning gpt-5.4-nano-2026-03-17 is not a way around that retirement. For a short-lived workflow, record the model ID and test date; for a longer-lived one, put the model behind configuration and test the recommended replacement with the same fixture.
Connect a local browser without handing over the session
Install Tabbit Browser and the Tabbit CLI skill. In Codex, use $tabbit to request the installed skill; its local launcher is ~/.local/bin/tabbit-cli. This is the skill syntax for Codex, not proof that GPT-5.4 nano is available as a Codex model choice. For this model, select an API-capable agent client whose current documentation explicitly supports OpenAI's gpt-5.4-nano plus local CLI or MCP tools. The client and model route need separate checks.
Confirm model availability: select the OpenAI API provider and exact ID
gpt-5.4-nano, then verify it in that account/client. A directory may displayopenai/gpt-5.4-nano; follow the selected provider's request format.Confirm skill discovery: ask the client to list or load
tabbit. Other clients use their own syntax; Pi uses/skill:tabbit, and some clients expose/skillsor discover skills automatically. Do not assume/tabbitworks everywhere.Confirm local reachability: run
tabbit-cli diagnoseand verify the intended Tabbit Browser instance. Name the task clearly and use a unique request ID for every CLI call.Confirm site access: open an approved support queue yourself in the target Tabbit environment. Check that the site is signed in and that this account may read the queue.
The browser page can reuse the login already active in that Tabbit environment. Tabbit CLI creates a separate task page so you can keep using your ordinary tabs. In the local fixture check, a protected page returned 401 before a short-lived test session was set; a second named task then read the page and returned the heading “Release notes review” with fixture record GEO-104. That verifies session reuse in the local fixture only. It did not connect GPT-5.4 nano, read support tickets, or test a client integration. The Tabbit CLI guide and existing-browser guide explain task pages and the existing session.
Extract only ticket IDs and links
Start with a support queue or ticket fixture that is approved for read-only testing. If the view contains customer names, emails, message text, attachments, or account IDs, do not include them in the model output. The task is an index, not ticket triage: return the visible ticket ID, the URL that opens that ticket, and a note that a person should verify the pair. Do not open reply, assign, close, or submit controls.
Use a prompt that sets the scope and the stopping point:
Task: create a local review draft named "GPT-5.4 nano ticket link index — M195".
Use Tabbit Browser through the installed Tabbit CLI skill. Read only the support queue and ticket pages I name, and only fields visible to my signed-in account.
For each requested ticket, return exactly: visible ticket ID, direct ticket URL, source queue URL, and review status. Do not copy names, email addresses, message text, attachments, or account identifiers. Do not infer a missing ID or build a URL from a guessed pattern.
Keep the result as a draft table. Stop before replying, assigning, changing status, closing, saving, or submitting anything. Show the source URL beside every ticket link so I can verify it.Illustrative output using synthetic values only; it is not a live support ticket or a result from the local fixture:
| Ticket ID | Ticket URL | Source queue | Review |
|---|---|---|---|
DEMO-SUP-104 | https://support.example.test/tickets/DEMO-SUP-104 | https://support.example.test/queue | Synthetic example; verify against the signed-in page |
This output format keeps the useful connection between a short identifier and the page where a reviewer can check it. It deliberately omits ticket content. The first browser-agent task checklist has a minimal way to verify that the client is reading the intended page; browser MCP vs CLI explains the trade-offs between the two transports.
Review each row before using it
| Field | Agent output | Human checkpoint |
|---|---|---|
| Ticket ID | Exact visible number, including prefixes and punctuation | Open the linked ticket and compare the ID character by character |
| Ticket URL | Direct URL read from the ticket page or its visible link | Confirm it opens the same ticket, not the queue or a neighboring record |
| Source queue | URL of the list from which the ticket was selected | Confirm the source view is authorized and current |
| Review status | “Needs human verification” | Keep it as a draft until each row is checked; do not act on uncertain rows |
The second check catches plausible but mismatched links. If a queue changes while the agent is reading it, start a fresh observation and verify the row again instead of trusting an old page snapshot. Never let the model manufacture a direct URL from a ticket-number pattern unless the page visibly provides that link. The guide to browser research with sources covers source-linked claims, and the browser-agent permissions guide discusses keeping actions within scope.
A Tabbit Browser workflow for support references
Tabbit Browser provides the signed-in page context; Tabbit CLI lets a local agent client inspect it through a separate task page. The existing login remains in the target Tabbit environment. You do not export a cookie or password to the prompt. The returned ticket ID and URL are still sent to GPT-5.4 nano when the client calls the hosted API, so keep the result to the minimum fields needed.
A safe sequence is: open the authorized queue in Tabbit, start a named CLI task, read the visible identifier and link, build the local draft table, and show it beside the source page. Then inspect each ID-link pair yourself. If your client is remote, it needs an explicitly supported local handoff to reach Tabbit CLI; an API model endpoint cannot see local tabs by itself. See local versus cloud browser agents and the Codex and Tabbit CLI setup guide for those boundaries.
The local session fixture is evidence for the browser connection in this environment, not a support-ticket test. No support site or ticket data was read for this guide. Treat the table above as an illustrative format and keep your first live run on a synthetic or explicitly approved queue.
Plan the model migration now
As of November 10, 2026, OpenAI's API docs label GPT-5.4 nano deprecated and list April 1, 2027 as its shutdown date. The recommended replacement is gpt-6-luna. That is a migration notice, not a promise that every client exposes the replacement or that its behavior is identical.
Keep the model ID in configuration rather than in workflow logic. Save a small, approved fixture with known ticket IDs and visible direct links. Before switching models, run both on that same fixture, compare exact ID-link pairs, note unknown or ambiguous rows, and inspect the provider/client logs. If the replacement is not yet available in your selected client, keep the workflow in draft and choose a currently supported route. The existing browser connection guide explains the session side; OpenAI's deprecation notice is the source of the API shutdown date.
GPT-5.4 nano is a reasonable short-term choice for a narrowly defined extraction task while its API route remains available and your client supports the needed local tool loop. Keep the browser task read-only, return only IDs and URLs, and leave the review checkpoint visible. Then test the documented replacement before the shutdown date rather than assuming the model name or behavior will carry over.
Related reading
FAQ
Can GPT-5.4 nano browse a signed-in website by itself?
No. GPT-5.4 nano supports function calling and MCP, but a client must provide and run the browser tool. Its OpenAI model page does not list the built-in computer-use tool for this model. A local client can still use an approved Tabbit CLI tool if it supports that handoff.
What is the exact GPT-5.4 nano API model ID?
OpenAI's API docs use `gpt-5.4-nano`; the catalog label `openai/gpt-5.4-nano` is a provider-qualified listing, not necessarily the request string for OpenAI's own API. Use the identifier documented by the endpoint you select.
Is GPT-5.4 nano available in ChatGPT?
OpenAI documents GPT-5.4 nano as an API model. Do not assume it can be selected directly in the ChatGPT model picker; verify current access in the API account and client you plan to use.
Is GPT-5.4 nano still supported?
As checked on November 10, 2026, OpenAI marks `gpt-5.4-nano` as deprecated and lists April 1, 2027 as its API shutdown date. OpenAI recommends `gpt-6-luna`; verify access and repeat your task evaluation before migrating.
Will GPT-5.4 nano see my browser password or ticket contents?
The browser session can remain in Tabbit, so you should not copy a password or cookie into the prompt. Any ticket fields returned to a hosted model are still sent to the API. For this task, return only the visible ticket ID and URL, and check the provider's data terms.
Can GPT-5.4 nano update or close tickets?
This workflow is read-only: collect IDs and links, then stop at a named review draft. Do not expose update, close, reply, or submit actions to the first test. A person should inspect the source and decide what happens next.