Comparing how two models handle the same prompt used to mean two browser tabs and no memory of which one said what. The redesigned Gateway Playground is a chat client built directly on your Gateway routes: conversation turns are bubbles you edit in place, a tool call and its result read as one guided step, and a single run streams in with a quiet line for tokens, latency, and cost. Compare mode lines up several models side by side in an arena view, with per-model colors, an optional blind-judging mode, and a latency photo finish; picking a winning reply carries it forward into the conversation.
Structured output follows the same rules as Logfire's Pydantic AI integration: native where a model supports it, a tool-mode fallback where it doesn't. When a session is ready to leave the Playground, export it as a full-fidelity Pydantic AI replay script, a Gateway curl command, or a runnable Logfire-instrumented starter that reads its API key from the environment rather than embedding it. An "Experiment in Playground" action on any agent run or Live view span preloads that exact exchange to keep iterating on.
Where to find it: AI Inference → Playground in your project. See the Playground guide.