Anthropic swapped its Workbench for a new Playground that stores no data on the provider’s servers, stripping out saved prompts, version history, evals, and team‑sharing features. OpenAI’s Playground, a legacy tool dating back to 2020, remains functional but will retire its saved‑prompt and eval services later this year. Both moves signal a shift toward treating prompts as part of source code rather than as artifacts in a web console, a change that directly affects how AI engineers, platform teams, and security practitioners build and operate prompt‑driven workflows.
What Changed in Anthropic’s Playground
The August 18 update replaced the Workbench UI with a leaner Playground interface. Key differences include:
- All prompt data is discarded after execution; the service is explicitly stateless.
- Features such as saved prompts, version history, evaluation suites, and team sharing were removed.
- Existing Workbench data must be exported before September 1, after which it will be inaccessible.
- The default model is
claude-sonnet-5, with no model‑switching UI for users without additional credits.
Exporting a session produces a ready‑to‑run Python script that embeds the original instructions and calls the model directly, requiring no post‑export edits.
Comparison with OpenAI’s Playground
OpenAI’s Playground (now labeled “Chat”) still offers a richer UI, including controls for reasoning, verbosity, and a button that auto‑generates prompts. However, the platform is also moving away from stored prompt artifacts, with a November 30 shutdown of its saved‑prompt and eval services.
During a head‑to‑head test building a PR‑review bot:
- Anthropic returned the correct JSON response in 1.9 seconds, using 102 tokens at a cost of $0.0029, and the exported script ran unchanged.
- OpenAI produced a correct answer in 6.2 seconds, consuming roughly 2.9 k tokens; cost was not displayed. The exported file contained the full session transcript, requiring manual removal of prior responses and insertion of a print statement before it could be executed.
When testing token‑limit handling, Anthropic clearly indicated a max‑token cut‑off with a plain message. OpenAI did not expose a comparable cap, preventing the same failure scenario from being reproduced.
Operational Implications of Stateless Prompt Testing
Because Anthropic’s Playground does not retain any data, teams lose built‑in versioning and collaborative sharing. Practitioners must now rely on external version control systems to track prompt revisions, which can improve auditability but adds process overhead.
Export behavior also diverges:
- Anthropic’s export is a self‑contained script, simplifying CI/CD integration; the script can be checked into a repository and invoked from automation pipelines without modification.
- OpenAI’s export bundles the interactive transcript, meaning additional scripting is required to isolate the model call before it can be used in automated contexts.
Both providers’ decisions to retire saved‑prompt services reinforce the principle that prompt definitions belong in code repositories, aligning with existing software‑delivery practices and reducing reliance on proprietary UI state.
Related CloudNinjas coverage: AI engineering.
What This Means For Practitioners
Teams should treat prompt engineering as code: store prompts in version‑controlled files, include them in CI pipelines, and avoid relying on web‑based storage for production workflows. When using Anthropic, plan for external collaboration tools to replace the removed sharing features. With OpenAI, be prepared to script out the relevant model call from exported transcripts before automation. Finally, monitor the upcoming deprecation dates (September 1 for Anthropic Workbench data, November 30 for OpenAI saved prompts) to ensure any lingering assets are migrated before they become inaccessible.
