AI playground

Open the console playground.

The playground lives under AI → Playground in the Cloud console sidebar, alongside AI → Settings. Sign in and use the project selector in the top bar. The console loads models automatically and uses a server-managed project key; there is no API key to paste or create for playground requests.

Owners can enable AI from Settings and submit billable work. Members can browse models and export requests, but cannot generate from the console. Use the tenant selector for a project's active tenant. Billing is charged to the selected project's linked account; the context panel links to Billing.

Choose Text, Images, Audio, Video or Embeddings first, then select a canonical model, its provider, and an API style in the settings sidebar. The main workspace changes into chat, an image canvas, an audio studio, or the selected task. The workspace selector shows the number of configured models. An empty task explains what needs configuration. Images selects native Google image generation or the OpenAI image endpoints; Audio includes native Google speech generation and configured audio endpoints. Live transcription has its own microphone panel. Review the JSON and constraints, then press Run. Run comparison submits two independent paid requests. Each gets its own stable idempotency key. Text streams, JSON, images, audio and video use their endpoint's response format. Refresh tasks for final billing. Exported curl, JavaScript and Python use an environment variable for your own scoped API credential.

The playground uses bounded buffers: uploads under 19 MiB, JSON/audio/SSE up to 32 MiB, and video previews up to 64 MiB. Use a server client for larger media. Closing the page or switching projects does not cancel an accepted provider job. A retry keeps its original request and idempotency key.

If AI is disabled or the reviewed catalog is empty, the playground shows that state. See errors.

Text conversations preserve up to twelve turns locally while you switch text API styles. Clear the conversation to start again; reloading the page or changing project/tenant clears it. Each subsequent turn includes the visible history and bills for its input. Image controls use the selected client schema: OpenAI image sizes or Google aspect ratios and resolution tiers. Parameter edits survive prompt edits. Advanced request JSON and exported code are expandable, keeping the workspace focused on content.

Google speech models expose voice selection and, where advertised, two speaker names, voices and dialogue fields. Live transcription is expandable within Audio. Capabilities come from the selected provider route. See API styles for translation boundaries.

Studio layout and model discovery

The model card in the right-hand Run settings panel opens a searchable catalog across model types and providers. Use category chips and the provider filter to narrow it. Selecting a model changes the workspace and shows only its reviewed settings. Provider selection lists routes serving that same canonical model; API style is selected separately.

System instructions appear only when the model advertises support. Image controls follow the selected API style. Speech models show voice and speaker controls; OpenAI speech supports one speaker, while reviewed native Google speech routes can support two. The bottom composer supports Command+Enter on macOS and Ctrl+Enter on other platforms. Get code opens the current request without exposing your project key.

The default catalog includes a snapshot of bounded OpenRouter text routes. These routes accept text input and settle from the exact account usage.cost receipt, including provider-specific cache charges. Published per-token tariffs do not replace the final receipt. No funds are held before submission. The snapshot expires after 30 days and must be refreshed and republished. Native media, live conversation, hosted agents and batch variants require separate reviewed transport and settlement support; listing a provider's model does not automatically enable those operations.

Image generation routes with reviewed reference support show an attachment tray in the composer. Use Add images or drag files onto the tray, preview and remove references, and review the route-specific count/size limits. Ratio controls display the corresponding shape; quality, resolution, provider and API style use option buttons. Selecting a direct OpenAI reference automatically changes to image edits. Google inline images and OpenRouter JSON references remain in the selected client style, and Get code exports those reference bytes.

Export the selected API request

Get code includes provider, API style, operation, and request mode selectors for the selected model. Only configured providers serving that canonical model and reviewed API styles for the current task are offered. These choices also update the playground. Google speech supports OpenAI speech where advertised; multi-speaker or different per-block delivery styles require Google style.

The read-only Monaco preview follows the current text, voice, speaker blocks, delivery instructions, conversation history, API style and advanced JSON, including text media attachments and image references. Changing the style rebuilds the payload. Choose Direct request in Get code to export a direct /v1/google/generateContent request; choose Durable task to export the /v1/ai/tasks envelope. The same choice controls Run. Native Google audio is base64 inside JSON; OpenAI speech returns binary bytes. See the speech examples for saving both kinds of result as WAV files.

Documentation examples are starting templates. To export your own project, tenant, prompt, and model settings, use Get code in the playground and copy the generated curl, JavaScript, or Python. Long snippets scroll inside the bounded preview; the dialog also scrolls on small screens.