Anthropic is shutting down the legacy Workbench and three experimental prompt tools APIs on August 17 — 15 days from today. If you have saved prompts, eval test cases, or variables stored in the legacy console, they do not carry over to the new Workbench. Export your data now, before it is gone permanently. And if any of your code calls /v1/experimental/generate_prompt
, /v1/experimental/improve_prompt
, or /v1/experimental/templatize_prompt
, those endpoints start returning errors on August 17.
What Is Being Retired #
Two things are going away simultaneously. First, the legacy Workbench — the older version of Anthropic’s browser-based prompt testing environment. New users have not had access to it for some time, but existing users still do, until August 17. Second, three experimental API endpoints that powered prompt generation and improvement:
/v1/experimental/generate_prompt
— described what you wanted, Claude generated a prompt template/v1/experimental/improve_prompt
— passed in a prompt, Claude rewrote it with chain-of-thought and specificity techniques/v1/experimental/templatize_prompt
— converted a concrete prompt into a reusable template with{{VARIABLE}}
placeholders
These were always labeled “experimental” — Anthropic’s way of signaling useful-but-not-committed. That label mattered less when the tools were working and free. It matters now that they are being removed without replacement.
Your Data Disappears Without Export #
This is the part that will catch teams off guard. The legacy Workbench stored your saved prompts, your variables, and your eval test cases. None of that migrates to the new Workbench. When the legacy version shuts down, it takes your data with it.
To export before August 17: open the Console, look for the export banner, or go to your Organizational Settings. If you have eval test suites built in the legacy Workbench, this is your highest-urgency item. Regression-testing infrastructure built on saved evals needs to move somewhere else entirely — there is no equivalent in the new Workbench.
How to Replace the Experimental Endpoints #
Anthropic is not providing replacement endpoints. The functionality is going away at the API level. If your codebase or CI/CD pipeline calls any of the three endpoints, you have two options: remove the calls, or replicate the behavior yourself using the Messages API directly.
Here is how to replicate improve_prompt
with a direct Messages API call:
response = client.messages.create(
model="claude-opus-5-20260801",
max_tokens=1024,
system="You are an expert prompt engineer. Improve the following prompt using chain-of-thought reasoning, clear role definition, and specificity. Return only the improved prompt.",
messages=[{
"role": "user",
"content": f"Improve this prompt: {your_prompt}"
}]
)
The improve-prompt behavior is still accessible in the new Workbench UI — just not via API. For any team that automated prompt improvement in a pipeline, the Messages API is the only path forward. Build the system prompt that instructs Claude to act as a prompt engineer, and you have rebuilt the feature.
The New Workbench: A Thinner Tool #
The replacement is deliberately simpler. The new Workbench is stateless — nothing stored on Anthropic’s servers. It is built directly on the public Messages API, meaning the request you build in the UI is the exact same request you would send in production code. A code tab exports any request as a Python or JavaScript snippet in one click.
What it does not have: saved prompts, variables, evals, or any API equivalent of the three endpoints being retired. Anthropic is positioning it as a testing and debugging tool, not a prompt management platform. Teams that relied on the legacy Workbench as a lightweight prompt store need to find a proper alternative — Git is the obvious default, or a dedicated platform like PromptLayer, LangSmith, or Braintrust if you need versioning and evals.
Anthropic’s Direction #
This is consistent with a broader platform shift. Anthropic is consolidating around the Messages API as the single interface. The new Workbench is a thin UI wrapper on that interface — not an abstraction above it. Proprietary prompt storage and experimental helper endpoints do not fit that model. What you get instead is a cleaner separation: the API handles inference, you handle prompt management.
Whether that trade-off works for your team depends on how embedded those experimental features were. For most teams, this is a minor inconvenience. For anyone running automated prompt pipelines against these endpoints, the clock is ticking.
Export your data. Audit your codebase for the three endpoints. August 17 is 15 days out.