Anthropic shuts down legacy Claude Workbench and three experimental API endpoints on 17 August 2026
The legacy Claude Workbench and three /v1/experimental/ prompt API endpoints reach end of life on 17 August 2026. Here's what to do.
17 August 2026 is the hard shutdown date for the legacy Claude Workbench and three experimental prompt API endpoints. Announced on 17 July 2026 with just 31 days’ notice, this retirement is now live. If your team has not yet exported its saved prompts or migrated away from the deprecated endpoints, that data and those integrations are gone.
Here is exactly what has been removed and what you should do about it.
What has been shut down
The legacy Workbench at platform.claude.com/workbench is no longer accessible. This is the original prompt-building interface inside the Claude Console where teams could save prompts, manage variables, and run evaluations. It is not the same as the refreshed Workbench at platform.claude.com/playground, which launched alongside this deprecation.
Critically, the new Workbench does not carry over anything from the old one. Saved prompts, prompt versions, and evals stored in the legacy tool do not migrate automatically. The only path was a manual export before the shutdown date.
Three experimental API endpoints have also been removed simultaneously:
/v1/experimental/generate_prompt/v1/experimental/improve_prompt/v1/experimental/templatize_prompt
Any code still calling these endpoints will now receive a hard error. There is no redirect, no grace period, and no named replacement endpoint for any of the three.
Who is affected
If your team started using the Workbench before 16 June 2026, you had access to the legacy version up until the shutdown date and should have seen a banner in the interface pointing you to the export option. If your account was created on or after 16 June 2026, you never had access to the legacy Workbench at all.
On the API side, if you built any internal tooling or automation that calls those three /v1/experimental/ endpoints, your integrations are now broken. These endpoints were designed for programmatic prompt generation, improvement, and templatising, and they have been removed without a direct successor.
What the new Workbench does differently
The refreshed Workbench at platform.claude.com/playground is built directly on the public Messages API. The request you build in the interface is exactly the request sent in code, which makes it easier to move from prototype to production. You can use the “code” tab to copy any request directly into your application.
The trade-off is that the new Workbench does not store prompts or conversations on Anthropic’s servers. Your current draft lives in the browser only. Anthropic’s guidance is to store exported prompts alongside your application code, in the same repository where you manage everything else. It is a deliberate architectural shift away from a cloud-hosted prompt library and towards treating prompts as code artefacts you own and version yourself.
What you should do now
If you missed the export window, the hard reality is that prompts stored only in the legacy Workbench are no longer accessible. There is no recovery path described in Anthropic’s documentation.
If you are still calling the deprecated endpoints, you need to remove or replace those calls immediately. Requests are now returning errors, and there is no fallback. Anthropic has not specified replacement endpoints for the three retired experimental routes, so you will need to implement equivalent functionality using the standard Messages API directly.
For your prompt library going forward, the recommended approach is to treat prompts as versioned text files stored in source control alongside your application code. The new Workbench gives you a scratchpad for building and testing, but persistence is your responsibility.
A word on the notice period
31 days is a short runway for a deprecation that removes a prompt library with no automatic migration. Anthropic has been accelerating its deprecation cadence across the platform, and this announcement sits alongside a broader set of upcoming retirements: Claude Sonnet 4.5 is tentatively scheduled for 29 September 2026, Claude Haiku 4.5 for 15 October 2026, and Claude Opus 4.5 for 24 November 2026. If you are building on the Claude platform, it is worth checking the model deprecations page regularly and treating any /v1/experimental/ endpoint as inherently short-lived.
The broader pattern here is clear: Anthropic is tightening up its API surface and moving away from hosted tooling that stores state on its servers. For teams building on Claude, that means owning more of your own infrastructure around prompts, evals, and version history, rather than relying on platform-side storage to hold it for you.