high. Use low, medium, high, or xhigh; minimal and thinking budgets are not supported.
The default output budget is 65,536 tokens. Set a larger budget for automatic continuation within the shared 1,048,576-token input and output window. Reasoning and tool calls count toward output.
Preserve all returned providerMetadata.google.parts across turns. Only the first function call in a parallel turn carries a thought signature. Return tool results in call order and end history with a user message or completed tool results.
For applications that compact or edit earlier history, set providerOptions.google.generationConfig.thinkingConfig.prefixMismatchBehavior to DROP; the default is ERROR.
Custom temperature, top_p, and top_k values are ignored. Frequency and presence penalties and multiple candidates are not supported. Implicit caching starts at 4,096 tokens; explicit caching is not supported.
Office and OpenDocument files can be supplied as Base64 data URLs with their correct MIME type through file content parts, including tool results.
Enable Google Search or URL Context through providerOptions.google.tools. Search is charged per actual query; search-result input tokens are free. Use Search separately from other built-in tools and Agentic video.
Video content parts accept mediaProcessing: STATIC (default) or AGENTIC. Count input tokens without a wallet charge through POST /v1/token-counts.
Model ID
model field.
Supported modes
chat_completionsresponses_apivisual_understandingtext_generation
Inputs
TextImageVideoAudio
Outputs
Text
Capabilities
- Not specified
Additional capabilities
Chat completionsFunction callingCode executionComputer useResponses APIVideo understandingSSE streamingMultimodal understandingReasoningStructured outputGoogle Search groundingURL ContextAgentic video understandingToken countingSee Function calling for tool declarations and result submission.
Technical details
Availability and final customer prices can change. Check the Dashboard model catalog before production use.