Anthropic’s Playground vs. OpenAI’s: The week-old tool beat the six-year incumbent

Anthropic swapped Workbench for Playground on August 18, and it didn’t just rename the tab. The company cut saved prompts, version history, evals, and team sharing, and the new tool keeps nothing on Anthropic’s servers. Anyone with old Workbench data has until September 1 to export it.

That timing matters because OpenAI made a similar move the same week, announcing that its saved Prompts and Evals platform will shut down on November 30. The message from both companies is blunt: prompts belong in code, not in a web console. The difference is in how well the replacements actually behave.

The New Stack tested both tools with the same prompt and the same code diff, building a tiny PR review bot that had to return JSON with risk, touched files, a summary, and whether tests changed. Anthropic’s Playground used claude-sonnet-5 by default, answered correctly on the first try, and exported cleanly into Python that ran in a terminal without edits. It finished in 1.9 seconds, used 102 tokens, and cost $0.0029.

OpenAI’s old Playground, now called Chat, also got the first answer right. It was slower at 6.2 seconds and offered extras Anthropic doesn’t, including color-coded formatting and a view of the model’s reasoning. But the export was a mess: the generated Python replayed the whole conversation, including the model’s answer, so running it produced nothing useful until the test was patched by hand.

The failure test was even worse for OpenAI. Anthropic clearly explained a max-token cutoff with a plain error message under the clipped response. In OpenAI’s Chat, the relevant setting couldn’t be found anywhere the tester looked, including the overflow menus and the older API mode behind a warning popup. On these tests, the newer, leaner tool won because it did the basic things right.