Anthropic publishes Claude Opus 4.8 as a distinct API model with ID claude-opus-4-8. It is a closed, hosted model, unlike open-weight Step 3.7 Flash. Opus 5 and Opus 5.5 are separate later choices, not aliases for 4.8. Anthropic’s lifecycle page currently lists Opus 4.8 as active; check it again before deployment. A provider or client may expose a different selector or entitlement, so verify the exact model where you will run it.
Key takeaways
The model proposes; the local client manages tools and permissions.
Tabbit Browser supplies the browser environment. Tabbit CLI exposes a task tab to a compatible local agent.
Check the exact model route and the actual browser account before the task.
Start read-only. Leave a named draft and check every cited page.
Pick the model route before opening the page
| Check | Claude Opus 4.8 | Browser task requirement |
|---|---|---|
| Model access | Confirm exact ID and provider in the client | Not provided by the browser |
| Execution | Provider-hosted inference; Anthropic hosts the model | Local agent must run the Tabbit CLI |
| Version boundary | Opus 4.8 is separate from Opus 5 and Opus 5.5 | Do not substitute versions silently |
| Login context | Not part of the model endpoint | Confirm the account in Tabbit Browser |
Anthropic publishes Claude Opus 4.8 as a distinct API model with ID claude-opus-4-8. It is a closed, hosted model, unlike open-weight Step 3.7 Flash. Opus 5 and Opus 5.5 are separate later choices, not aliases for 4.8. Anthropic’s lifecycle page currently lists Opus 4.8 as active; check it again before deployment. A provider or client may expose a different selector or entitlement, so verify the exact model where you will run it.
A safe first task: compare a dashboard record
Choose an approved, low-risk record with a stable ID. Confirm it in the Tabbit Browser yourself, then ask the local agent to inspect the visible status and updated date. The goal is a comparison note, not a site change. If the record ID differs, a field is missing, or a login prompt appears, the task should stop and report that boundary.
| Checkpoint | Keep in the draft | Stop for human review when |
|---|---|---|
| Record identity | Stable ID and page URL | ID is missing or differs |
| Each field | Visible value plus source URL | Value is inferred or stale |
| Result | Match, mismatch, or unknown | Page asks to save or submit |
| Next action | Named review note | Any change to the source system |
A prompt for a named review draft
Use this prompt with your own test record. It is a suggested workflow, not a report of a model benchmark or a completed test against a third-party account.
Task name: GEO-104 dashboard change review
Use Claude Opus 4.8 in this local agent and the installed Tabbit skill. Open the two URLs I provide in a separate Tabbit task tab. Read only record ID, status, and updated date. For every value, include its source URL. Use unknown when a field is not visible. If the ID differs, a login screen or CAPTCHA appears, or the page offers a write action, stop and report it. Do not save, submit, send, or change anything. Leave a draft named “GEO-104 — dashboard change review” with matched fields, mismatches, unknowns, and the URLs for my review.Use the model you can actually select in a local agent client. Then connect that client to Tabbit CLI. The model does not supply the browser, login session, or permission boundary. Start with read-only inspection and leave a named draft for a person to check.
A practical setup
A Reddit user described using Opus 4.8 across several coding projects and called it slower than Opus 5. Another thread debates whether to choose 4.8 or 5, with strongly subjective claims about instruction following. These posts show that model preference depends on workload; they do not establish browser task performance. Read the project-use report and the version discussion. Anthropic’s model page and lifecycle documentation describe the model and its API status. Claude Code or another client may expose model selection, but do not assume it can run Tabbit CLI until you check its local skill and launcher permissions.
For the model overview, lifecycle, pricing, and feature context, see the published Claude Opus 4.8 overview. For setup, see the Tabbit CLI guide, the existing-browser connection guide, the first-task checklist, and the CLI versus MCP guide. The Codex setup example and login reuse guide cover adjacent setup details. Use the browser prompt template to make stop conditions explicit. If an action is unavailable, use the click troubleshooting guide. For model-side reasoning background, see agentic reasoning in LLMs.
First confirm the account and target record in Tabbit Browser. In a compatible local client, discover the installed Tabbit skill. Clients with named skill commands may use /tabbit; Pi uses /skill:tabbit; some clients expose /skills. Skill discovery, launcher reachability, model selection, and site login are separate checks. Run the installed tabbit-cli diagnose, then start a uniquely named task. The CLI operates in the Tabbit environment, where a task can use the existing signed-in session. A separate task tab lets you keep browsing while it works. This reuse was checked with a local test fixture, not with this model/client combination.
Draft a named review
Anthropic’s published API ID and model lifecycle page are the source of truth for current availability. A model being active does not mean every third-party client, plan, or account can select it.
FAQ
What API model ID should I use?
Anthropic documents `claude-opus-4-8`. Confirm your account and chosen provider expose that exact ID before running the task.
Is Opus 4.8 the same as Opus 5 or 5.5?
No. Each is a separate model. Keep the exact version in your task configuration and notes so you can reproduce the result.
Is Opus 4.8 open source?
No. It is a hosted Anthropic model. The model weights are not published for local deployment.
Does the API give it access to my browser?
No. Browser access comes from a local agent and its connector. Tabbit CLI connects a compatible client to a Tabbit browser task tab.
How can I leave work for review?
Ask for a named draft with source URLs and unknowns called out. Keep the task read-only and verify the visible pages before taking action.
Could the model be retired?
Model status changes. Anthropic’s lifecycle page lists current status and dates; check it before building a workflow that depends on this ID.