Anthropic Playground is the name the company now uses for the prompt-testing surface in the Claude Console, and the change is considerably more than cosmetic. The official release note is one sentence — “Workbench is now playground in the Claude Console” — but the product underneath it lost saved prompts, version history, variables, evals and sharing, and gained something the old tool never had: the full SDK request for every run, shown next to the API response.
The dates matter because one of them has already passed. Legacy Workbench access ended on 17 August 2026. Playground went live in the Console on 18 August 2026. The export window for anything still sitting in Workbench closed on 1 September 2026, and Anthropic’s help centre is blunt about what that means: “This data will no longer be recoverable after September 1, 2026.”
If you are reading this after that date with prompts you never exported, they are gone. This article covers exactly what changed, what the Anthropic Playground does and does not do, which API endpoints were retired alongside the old tool, and how to rebuild the workflows that depended on the features that went away.
Table of contents
- What the Anthropic Playground Replaced
- What the Anthropic Playground Actually Does
- Why a Stateless Anthropic Playground Is a Defensible Choice
- Rebuilding the Workflows the Anthropic Playground Dropped
- The Anthropic Playground Timeline in One Place
- What the Anthropic Playground Says About the Console’s Direction
- Who Is Affected by the Anthropic Playground Change
- Where the Anthropic Playground Sits Among Vendor Consoles
- Practical Notes Before You Open the Anthropic Playground
- Frequently Asked Questions About the Anthropic Playground
- References
What the Anthropic Playground Replaced
The old tool had been part of the Console since the earliest days of the Claude API, and it accumulated features accordingly.
Workbench was stateful by design
Legacy Workbench, at platform.claude.com/workbench, stored your work on Anthropic’s servers. You could save a prompt, come back to it, version it, share it with a colleague, define variables and attach evaluations. For a lot of teams it was the first place a prompt existed before anyone wrote code around it.
The sunset was announced a month ahead
The 17 July 2026 release note gave the schedule: access to legacy Workbench would end on 17 August 2026, and “saved prompts, variables, and evals are not supported in the updated Workbench.” Users were pointed at an in-product banner and Organizational Settings to export anything they wanted to keep.
The Anthropic Playground is stateless
The replacement keeps nothing. Anthropic’s help centre describes it as “a simpler, stateless way to try Claude models” built directly on the public Messages API. No stored conversation history, no saved prompt history or evaluations, no prompt sharing and no versioning. Drafts live in the browser and go away with the tab.
Three API endpoints went with it
The retirement was not limited to a web page. The experimental prompt tools APIs — /v1/experimental/generate_prompt, /v1/experimental/improve_prompt and /v1/experimental/templatize_prompt — were retired alongside Workbench on 17 August 2026. Requests to those endpoints now return an error. Any pipeline that called them to generate or refine prompts programmatically broke on that date.
| Capability | Legacy Workbench | Anthropic Playground |
|---|---|---|
| Saved prompts | Yes, server-side | No |
| Prompt history and versioning | Yes | No |
| Variables | Yes | No |
| Evaluations | Yes | No |
| Prompt sharing | Yes | No |
| Full SDK request shown per run | No | Yes |
| Every Messages API parameter | Partial | Yes |
| Feature templates (code execution, web search) | No | Yes |
| Export current request as code | Limited | Yes |
| Data stored on Anthropic servers | Yes | No |
What the Anthropic Playground Actually Does
Read the trade the other way round and the design intent is clear.
Every Messages API parameter is exposed
The release note is specific: Playground “supports every Messages API parameter.” That is a meaningful superset of what the old interface let you touch. Anything you can send in a Messages request — tool definitions, structured outputs, thinking configuration, system messages, sampling controls — you can set in the browser and run.
The request is shown, not hidden
“It shows the full SDK request and the API response for each run, to help you understand the API and build with it.” This is the philosophical shift. Workbench was a prompt editor that happened to call an API. The Anthropic Playground is an API client that happens to have a text box, and it teaches the API by showing it to you.
Templates demonstrate server-side features
It ships with templates covering API features such as code execution and web search — the server-side tools that are awkward to wire up from scratch and easy to get subtly wrong. A working example you can run and then export is a faster path than reading the reference.
Export to code is the intended exit
Once a request does what you want, you export it as a code snippet and it becomes part of your application. The tool is explicitly a staging post rather than a destination, which is the honest justification for making it stateless.
Where to find it
The Anthropic Playground lives at platform.claude.com/playground inside the Claude Console, and is covered in the Claude Help Center article “How do I use the playground?”, last updated 18 August 2026.
Why a Stateless Anthropic Playground Is a Defensible Choice
The reaction to losing saved prompts was not universally warm, and the reasoning behind it is worth setting out fairly.
Prompt storage was solving the wrong problem
A prompt that matters lives in version control alongside the code that sends it. A prompt stored in a vendor’s web tool is a prompt that is not in your repository, not in code review, not in your deployment pipeline and not diffable. The old workflow quietly encouraged the least reproducible way to manage a critical asset.
Evals moved up the stack
Anthropic’s platform has grown a great deal since the Workbench evals feature was designed, and serious evaluation now belongs in a harness that runs on every change rather than in a browser tab. Losing a lightweight in-console eval is a real cost for quick checks; it is not a gap in the platform.
Statelessness has a compliance dividend
Nothing typed into the Anthropic Playground is stored on Anthropic’s servers. For teams who were previously careful about what could be pasted into a hosted prompt editor, the answer is now structural rather than procedural.
The counter-argument is real
None of that helps someone who used Workbench exactly as intended, kept a library of tested prompts there, and had one month’s notice to move them. The transition cost fell entirely on users who trusted the tool’s persistence, and the migration path was an export button rather than an import target.
Rebuilding the Workflows the Anthropic Playground Dropped
Each removed feature has a replacement. None of them is in the browser.
Saved prompts go in the repository
Keep prompts as files next to the code that sends them — a directory of text or template files, or typed constants in your application language. You get diffs, review, blame and rollback, which the old saved-prompt list never offered.
Versioning is what git already does
A prompt change that alters behaviour is a code change. Treat it like one: branch, review, merge. Tagging the commit that shipped a prompt is more useful than a version number inside a vendor tool, because it ties the prompt to the model version and the surrounding code that were live at the same time.
Variables become template rendering
Workbench variables were string substitution. Any templating library in your stack does the same thing, testably, in the language your application already speaks.
Evals belong in a test harness
Build a small set of cases with expected properties, run them in CI against a pinned model, and record the results. This is more work than the old in-console evals and it is the only version of evaluation that catches a regression before a user does.
Sharing becomes a pull request
The thing a colleague needs is rarely the prompt alone — it is the prompt, the parameters, the model and the surrounding code. The Anthropic Playground’s export-to-code button plus a branch delivers all four.
| Lost feature | Where it goes now | Net effect |
|---|---|---|
| Saved prompts | Files in your repository | Reviewable and diffable |
| Versioning | Git history and tags | Tied to model and code version |
| Variables | Template rendering in your language | Testable, no new tool |
| Evals | CI test harness against a pinned model | Catches regressions automatically |
| Sharing | Exported code in a pull request | Shares parameters and model too |
| Prompt generation endpoints | Ordinary Messages API calls | No special endpoint to depend on |
The Anthropic Playground Timeline in One Place
Four dates cover the entire transition, and they are close enough together that missing one was easy.
| Date | What happened | Action it required |
|---|---|---|
| 17 July 2026 | Sunset announced in the release notes | Start exporting; audit calls to the experimental endpoints |
| 17 August 2026 | Legacy Workbench access ends; three experimental prompt endpoints retired | Pipelines calling those endpoints had to be changed |
| 18 August 2026 | Playground live in the Claude Console | Re-learn the interface; help centre article updated |
| 1 September 2026 | Export window closes permanently | Last chance; data not recoverable afterwards |
The gap that caught people out
Access to the old tool ended two weeks before the export deadline. Anyone who assumed the export lived inside Workbench, and went looking for it on 18 August, found the interface already replaced. The export controls sat under Organizational Settings, which is not where most people looked.
What the export actually contained
Anthropic’s help centre lists the export as covering saved work from the Console, with model completions and uploaded files as optional inclusions — and warns that adding them “may significantly increase the export size.” That warning is worth reading as a hint about what a large team’s Workbench history looked like by August 2026: not a handful of prompts, but years of accumulated runs.
Why the deadline was firm
A stateless replacement cannot hold the old data, and the announcement gave a specific, repeated date rather than a vague deprecation. Anthropic’s release notes and help centre both carried it, which is a better standard of notice than many retirements get — and still leaves no recourse for anyone who read neither. Teams tracking vendor changes across several providers at once will find our coverage of autonomous AI agents and the platforms behind them a useful place to start when auditing what else has a sunset date attached.
What the Anthropic Playground Says About the Console's Direction
One renamed page is a small thing. Read alongside everything else Anthropic shipped through 2026, it is part of a consistent pattern.
The Console is becoming an operations surface
The same release-note stream that carried the Playground change also carried a redesigned session viewer with a timeline minimap, a transcript grouped by model request, and an inspector panel for cost, raw events, per-tool statistics and mounted resources. That is observability tooling for systems already in production, not authoring tooling for prototypes.
Authoring is expected to happen in your own tools
Between an export-to-code button, a command-line interface that manages resources from files in a repository, and the removal of server-side prompt storage, the message is that the artefacts belong on your side of the line. The Console shows you what the platform is doing; your repository decides what it should do.
Statelessness is consistent with the rest of the API
The Messages API has always been stateless — you send the whole conversation each time. A prompt editor that quietly kept state on the server was the odd one out, and the Anthropic Playground now behaves the way the API it fronts behaves.
The trade-off is a higher floor for beginners
The honest cost of this direction is that the easiest possible on-ramp got steeper. Someone with no repository, no CI and no templating library had a usable workflow in Workbench and now needs three tools to reproduce it. The Anthropic Playground is better for people who already have infrastructure and worse for people who do not.
Who Is Affected by the Anthropic Playground Change
The impact is uneven, and it lands hardest on the people least likely to have noticed the release note.
Engineers building on the API
Largely better off. Full parameter coverage, a visible request, working templates for server-side tools and one-click export to code are all improvements over a prompt editor that abstracted the API away.
Prompt engineers without a repository
The most affected group. A workflow built on a stored, versioned, shareable prompt library inside the Console no longer has a home in the Console, and the replacement requires tooling and habits that may not exist on the team yet.
Teams with prompt-tooling pipelines
Anyone calling the three experimental endpoints had a hard break on 17 August 2026. If a service still calls them, it is returning errors now, and the fix is to implement generation or refinement as ordinary Messages API calls against a current model.
Anyone who missed the export
There is no recovery path. The help centre states plainly that data is not recoverable after 1 September 2026, and export options included model completions and uploaded files — which Anthropic warned would significantly increase the export size.
Where the Anthropic Playground Sits Among Vendor Consoles
Every major model provider ships a browser tool for trying prompts, and they have been converging on different answers to the same question.
The two design philosophies
One approach treats the console as a lightweight IDE: store prompts, version them, attach datasets and evaluations, and let a team collaborate inside the vendor’s product. The other treats it as a request inspector: expose every parameter, show the wire format, and hand you code to take away. The Anthropic Playground moved decisively from the first camp to the second.
What you gain by moving to the second camp
Transparency is the main one. A tool that shows you the exact SDK request removes an entire class of confusion about what the abstraction was doing on your behalf — the kind that surfaces later as a bug when the browser and the application behave differently for reasons nobody can see.
What you give up
Collaboration inside the product, and a low-friction place for non-engineers to work. A product manager who wants to try a wording change now needs either a developer or a repository, where previously they needed a login. That is a real organisational cost, even if the engineering argument for the change is sound.
The practical answer for most teams
Use the Anthropic Playground for what it is good at — exploring parameters, learning server-side tools, and producing a correct code snippet — and keep the prompt library, the version history and the evaluation suite in the same repository as the application. That split is what the tool is now designed to assume, and fighting it produces the worst of both.
Practical Notes Before You Open the Anthropic Playground
A few things are worth knowing before the first session.
Pin the model you are testing
A prompt that behaves one way on one model can behave differently on the next release. Note the exact model identifier in the exported snippet, and treat a model change as a reason to re-run your cases. Anthropic’s release notes are the authoritative record of what changed and when.
Watch what you paste
Statelessness protects what is stored, not what is sent. The request still goes to the API, and your organisation’s usual rules about customer data in prompts apply unchanged.
Use the templates as documentation
For anything involving server-side tools, starting from the shipped template and modifying it is faster and less error-prone than assembling a tool definition from the reference. This is where the Anthropic Playground is most obviously better than what it replaced.
Keep the browser tab in mind
Work exists only until the tab closes. Export anything you care about before you navigate away, because there is no draft recovery and no history to fall back on. Teams standardising on a set of AI models and tools should build the export step into the habit rather than relying on the tool to remember.
Frequently Asked Questions About the Anthropic Playground
Is Playground the same thing as Workbench renamed?
No. The name changed and so did the product. Playground is stateless and built on the public Messages API; Workbench stored prompts, versions, variables and evals on Anthropic’s servers.
Can I still get my Workbench data?
No. The export deadline was 1 September 2026, and Anthropic states the data is not recoverable after that date.
Where is the Anthropic Playground?
At platform.claude.com/playground, inside the Claude Console. It requires a Console login.
Did the API itself change?
Not as part of this. Playground is built on the same public Messages API available to any client. The separate change was the retirement of the three experimental prompt tools endpoints on 17 August 2026.
Does Playground support tools and structured outputs?
Yes. It supports every Messages API parameter, including tool use and structured outputs, and ships templates for server-side features such as code execution and web search.
Is there still a way to share a prompt with a colleague?
Not inside the tool. Export the request as code and share that, which carries the parameters and model along with the prompt text. Anyone comparing assistants across vendors may also find our coverage of rival AI agent behaviour and monitoring useful context.
