To use Claude Fable 5 with a signed-in browser, select the exact claude-fable-5 model in a local agent client, confirm the Tabbit skill is available, and have Tabbit CLI read one record from two pages in a profile you chose. Ask for a named comparison draft and stop before changing either site. Anthropic lists Fable 5 as an active legacy API model; Claude Code's fable alias usually selects Fable 5.1, so check the actual model ID before you start. Anthropic's Fable 5 model page and Claude Code model configuration document those separate model names.
This guide uses a concrete case: compare a support record in a CRM with its matching entry in a billing portal. Match on a stable case ID, return the status and last-updated time from each page, and prepare Fable 5 account check — CASE-4821 for a person to review. The record and draft name are illustrative fixtures, not a test performed in either service.
One Reddit author who repeated a business web-app task concluded, “For a job you repeat, I'd let the agent work it out once, then run the script and bring the agent back when it breaks.” That is a useful boundary for this example: use the model to resolve an unfamiliar cross-site task, then keep the result reviewable. It is one author's report, not evidence about Fable 5 or Tabbit. Read the original discussion.
A second Reddit project author described using the model “mainly for planning” before local TypeScript and Playwright perform the browser actions. That account points to a practical distinction between the model and the browser executor. It concerns a Fable 5.1 project, not this Fable 5 workflow. Read the project thread.
Key takeaways
The exact Anthropic API model ID is
claude-fable-5. The newer Fable 5.1 has a different ID.A model being active does not mean every provider account can select it. Check the model name in the client.
Claude Code and Tabbit CLI have different jobs. Claude Code supplies the agent; the local Tabbit skill connects it to Tabbit Browser.
A local browser task can use the chosen Tabbit profile's current website session. Confirm the profile and account yourself.
Start with a two-page read, stable IDs, source URLs, and a named draft. Review before any write.
Confirm the exact model and local browser route
Anthropic's model reference lists claude-fable-5 as active legacy, with a 1 million token context window and 128,000 token maximum output. The current Fable model is Fable 5.1. Both entries have separate IDs, and Anthropic's deprecation list says Fable 5 will not retire before June 9, 2027. Those lifecycle facts can change, so check the current model page before setting up a new workflow.
Claude Code's fable alias normally resolves to Fable 5.1. If you specifically need Fable 5, select the full name claude-fable-5 in the model picker or with /model claude-fable-5, then confirm the selected model in the session. A gateway or organization allowlist can reject that ID. A visible model choice is the evidence that matters, not a tutorial's old screenshot. Anthropic's model configuration guide explains aliases, full names, and provider-specific differences. For Fable 5's general profile, see the model overview, which covers the latest Fable release and links back to this version boundary.
The browser path is a separate check. Tabbit Browser runs the website session; Tabbit CLI gives a local client access to that browser; Claude Code supplies the model and decides which local tool instructions to follow. This workflow does not mean Tabbit provides Fable 5 or that Anthropic has a native Tabbit integration. See the Claude Code and Tabbit CLI setup for the client connection, and use the Tabbit CLI guide if the launcher cannot reach the intended browser instance.
| Question | What to verify | What it means |
|---|---|---|
| Is Fable 5 selectable? | The client shows claude-fable-5, not only fable or claude-fable-5-1. | The model route is available to this account and provider. |
| Can the client find Tabbit? | A fresh Claude Code session lists /tabbit after the skill is installed and Claude Code is restarted. | The client can discover the skill instructions. |
| Can the local tool reach the browser? | The Tabbit CLI diagnostic names the expected Tabbit instance. | The launcher can contact that local browser. |
| Is the website session correct? | The task tab shows the expected domain, account, and record. | The browser profile is signed in to the page you intended. |
These checks do not substitute for each other. The model picker cannot establish that a local CLI is reachable. A successful CLI diagnostic cannot establish that the right site is signed in. If you need to choose a browser connection method, compare browser MCP and CLI; for a session that should use an already-open profile, read how an agent connects to an existing browser.
Prepare a two-site comparison
Install Tabbit Browser and the Tabbit CLI using the official Tabbit CLI quick start. In Claude Code, install the Tabbit skill where that client discovers skills, restart Claude Code after installation, and type / in a fresh session to confirm /tabbit appears. Use the command documented by the client; do not guess a slash command if your agent uses another skill interface. The Claude Code permissions guide covers approval of the local shell command.
Before sending page content to a hosted model, check your organization's data rules. The browser can run locally while the model still receives text extracted from the page. Anthropic also documents prompt-injection risks for computer use. Keep the task on approved sites and stop for a person before consequential actions. The browser first-task checklist gives a small verification sequence; Claude's Playwright MCP setup and Claude Agent SDK browser tools explain other routes when a local CLI skill is not the right fit.
Ask the account owner to open the correct CRM and billing portal in the intended Tabbit profile. If the profile is already signed in, the local task can use that session. If it is not, sign in yourself and check the account before continuing. Keep the two pages in a dedicated task context so you can keep working in other browser tabs. Don't ask the agent to handle passwords, one-time codes, or account recovery.
| Stage | Agent's task | Evidence to inspect before continuing |
|---|---|---|
| Select | Pin claude-fable-5 and confirm the client reports that model. | Model name in the current session. |
| Open | Read the two exact URLs in a Tabbit task tab. | Domain, page title, account label, and record ID. |
| Match | Find CASE-4821 on both pages and collect status and updated time. | Each value is tied to its source URL; missing values stay unknown. |
| Draft | Produce Fable 5 account check — CASE-4821 as a local review draft. | The draft contains both links and marks any mismatch clearly. |
| Stop | Wait for the reviewer. | No edits, submissions, messages, or account changes have occurred. |
This sequence is an example workflow, not a claim that Fable 5 was tested against a CRM or billing portal. For other agents, the browser control tutorial covers the first read. If the task requires screenshots or deterministic selectors instead, compare Claude in Chrome and Playwright MCP.
Use this prompt for the first read
Replace the example URLs and record ID with pages you are authorized to read. Keep the request narrow enough to verify against the open tabs.
Use the Tabbit skill available to this local agent. In a separate task tab, open only these two URLs:
1. [CRM record URL]
2. [Billing portal record URL]
Compare record CASE-4821 by its stable ID. Return one table with these columns:
source page, page title, record ID, status, last-updated time, source URL.
Do not infer missing fields. Mark them unknown and report conflicting values separately.
Then prepare a local review draft titled “Fable 5 account check — CASE-4821” with a short summary and the two source links.
Do not edit either site, submit a form, send a message, download data, or change account settings.
Stop and wait for my review. If the model, skill, browser, or sign-in check fails, name the failed layer and stop.Expected output: two source rows for the same stable ID, any mismatch called out, and a named draft that links to the pages. If the ID appears on only one page, report one match and one missing record instead of guessing. A successful page load is not enough; check the visible account and the actual record. The browser comparison guide discusses how to validate the selected tool against a task rather than its label.
Review the draft before any write
Check the ID and status in both source pages, compare the timestamps, and open each source link yourself. If the sites disagree, keep both values and ask the responsible person which source controls. Do not ask Fable to silently reconcile or update records.
A summary is a new artifact. It is not permission to alter the CRM, billing portal, or a customer account. Treat an edit, approval, payment, message, or submission as a separate task with a separate human checkpoint. If you later repeat a stable workflow, first test its failure handling and keep a person in the loop when the page changes. The browser MCP versus CLI test helps decide whether to keep an agent in the loop or use a more deterministic route.
FAQ
What model ID selects Claude Fable 5?
The Anthropic API model ID is `claude-fable-5`. Claude Code's `fable` alias usually resolves to Fable 5.1, so pin the full model name and confirm the selected model in your client.
Is Claude Fable 5 still available?
Anthropic currently lists `claude-fable-5` as active legacy, with retirement no sooner than June 9, 2027. Your account and provider still need to allow that exact model.
How is Claude Fable 5 different from Fable 5.1?
Fable 5.1 is the newer current model and Claude Code's usual `fable` alias target. Fable 5 has a separate model ID and remains active as a legacy option, so choose it only when you need that exact version.
Does Tabbit include Claude Fable 5?
No. Claude Code or another client supplies the model, while Tabbit CLI connects that local client to Tabbit Browser. Check model access and browser access as separate steps.
Can Fable 5 use my signed-in website session?
A local Tabbit CLI task can read pages in the Tabbit profile it connects to. Open and verify the intended profile yourself first, and share page contents with a hosted model only when your data policy allows it.
Should the agent update the source record?
Keep the first task read-only and ask for a named draft with links to both source pages. Review the identifiers and differences yourself, then approve any separate edit or submission.