Playground
Last updated September 15, 2026
Trying models directly in the console — multiple protocol entry points, side-by-side comparison, a parameter panel, and image and video generation.
The playground is for trying models without writing code: comparing outputs across models, checking parameter values, or confirming how a model behaves under a particular protocol.
Billing and attribution
Playground calls are billed normally, exactly as calls made through the API.
Usage is recorded against your own playground key in the current workspace. That key is platform-managed, does not appear in the key list, and cannot be used for external calls. Logs and usage can still be attributed to the individual member who made the call.
Text conversations
Pick a model and start. Each conversation remembers the model it last used, so switching conversations does not disturb the others.
Side-by-side comparison
The same question can be sent to several models at once, up to ten panes side by side. Each pane picks its own model, provider, and parameters, and panes can be dragged to reorder.
This is the fastest way to compare models — same prompt, same moment, same parameters, so the difference comes from the model alone.
Entry protocol
You can choose which protocol a call enters through, to compare how one model behaves across protocols.
The parameter list changes with the protocol, because each protocol can express a different set of parameters.
Parameter panel
The panel lists only the parameters that model genuinely supports through the selected protocol. Unchecked parameters are not sent at all, leaving the provider's own defaults in effect.
Common parameters include system prompt, reasoning effort, temperature, top P, top K, maximum output tokens, presence and frequency penalty, verbosity, seed, and stop sequences.
This list comes from the parameter support matrix the platform maintains. A parameter missing from the list will not take effect for that model through that protocol — sending it anyway gets it dropped or rejected. See protocol conversion compatibility.
Routing
You can set a sorting strategy, or name a provider directly.
⚠️ Naming a provider bypasses routing entirely: if it is unavailable the call fails outright, with no fallback to another provider.
Reasoning
For reasoning models, the thinking phase is shown separately along with how long it took.
Image generation
With an image model selected, you can set aspect ratio, size, resolution, quality, number of images, output format, background, compression, moderation strength, and input fidelity — which options are available depends on the model.
Generated images can be previewed in the conversation and opened at full size.
Video generation
Video models support several input modes, depending on the model:
| Mode | Notes |
|---|---|
| Text only | Generated from the prompt alone, with no material |
| First frame | One image as the first frame, with the prompt describing the motion that follows |
| First and last frame | Upload both to fix how the shot starts and ends |
| Reference images | Up to 9, to keep the subject and style consistent |
| Omni reference | Upload images, video, audio, documents, or links and refer to them with @ in the prompt |
Configurable parameters include resolution, duration, aspect ratio, audio, prompt expansion, thinking before generation, watermark, and negative prompt.
Duration can be set to "automatic", letting the model choose based on the material and prompt. Aspect ratio can be set to "adaptive", following the input material.
Video generation is asynchronous and usually takes several minutes. You can leave the page and return to check progress.
Attachments
Attachments can be added to a conversation; the supported types and size limits depend on the model. When a model cannot accept a file type it says so directly rather than dropping it silently.
Managing conversations
Conversations can be renamed and deleted, and a single turn can be deleted as well (the question and answer go together, and cannot be recovered).
⚠️ Playground conversations are stored under your account so that you can pick them up after a refresh or later. For content you would rather not have retained, call the API directly instead. See data retention and privacy controls.
Default model
The model preselected for new conversations can be configured:
| Where | Scope |
|---|---|
| Personal settings | Your personal workspace |
| Organization settings | Every workspace under that organization |
Leave it empty to follow the platform default. Existing conversations each remember their own last model, so this setting affects new conversations only.
Organization settings can be changed by the organization owner and admins only.
Related
- Model catalog and model IDs — available models and capabilities
- Routing and failover — what pinning a provider implies
- Call logs — playground calls appear in the logs too