python-genai

Adds the Gemini 3.8 Flash model to the Python SDK and updates Flash model descriptions.

Added video understanding support to the Interactions API, along with streaming downloads to disk via client.files.download, audio/webm MIME support, and translation_config for GEAP. Also exposed ProcessingCallStep and ProcessingResultStep in the Interactions SDK.

Adds AudioTranscriptionConfigMode for audio transcription configuration, environment files support with a scotty file download helper, and TYPE_JPEG2000 to the VideoContent.MimeType enum.

The Gemini Python SDK v2.19.0 adds a mode enum (VERBATIM, SMART) to audio transcription configs, along with new capabilities like video resolution and extension parameters, step count for reinforcement tuning, and an IDLE state for live connections. The REQUIRES_ACTION live connection status is now deprecated.

Adds support for the gemini-3.7-flash model.

The deferred service tier is now publicly available on Vertex. Speech_config becomes a structured object, and model validators now build on first use instead of at import to speed up import times.

Adds the Gemini Robotics ER 2 Preview model, a TOO_MANY_TOOL_CALLS value to the FinishReason enum, and a top-level errors array to the Interaction resource (iAPI). Also auto-populates propertyOrdering for ResponseSchema and ResponseJsonSchema in the .NET SDK.

Adds an environment resource wired into the SDK, and Google Maps Tool grounding_types now support places and routing. Also communicates upcoming changes in the next major version.

GenerateContentConfig now supports GenerationConfig.audio_transcription_config and Part.audio_transcription, with a flat language_codes field on AudioTranscriptionConfig. API keys can now be used with project and location for enterprise mode.

Added audio_transcription_config to GenerationConfig and Part.audio_transcription support. Deprecated the Imagen generate_images, edit_images, and generate_videos methods (with prompt/text/image args), plus LiveConnectConfig.GenerationConfig, ahead of the next major version.

A new custom_vocabulary field for the Live API and a model selector are now available. ASR fields are now public in v1main and Interactions, and mTLS support is added for custom clients using google auth mtls.

Fixed a bug where Interactions inside triggers were not being normalized properly.

Adds a new Triggers resource to the GoogleGenAI client, AntigravityAgentConfig and CodeMenderAgentConfig to OpenAPI schema overlays, and per-modality prompt token counts in embedding responses for gemini-embedding-2.

Added environment_id to network config, response_format and translation_config to GenerationConfig, and Tool.exa_ai_search for Gemini Enterprise API. Also added retrieval result step and new fields to maps and search results, enabled function declarations during Agent Creation, and fixed event loop mismatch and Python 3.14 built-in shadowing issues.

Adds disabled_safety_policies support for computer use on both Interactions and Models, video generation and response format parameters, usage fields for steps, and MCP support for async generate_content_stream.

The interactions implementation has been completely replaced with no public API surface change. Adds audioOffset to VoiceActivity, interimInputTranscription to LiveServerContent, Computer Use API fields, and ServiceTier to UsageMetadata.

Adds audioOffset and interimInputTranscription to Live API components, plus LanguageAuto, LanguageHints, and adaptationPhrases to AudioTranscriptionConfig. Also maps gemini-3-flash-preview to the local tokenizer and broadens publisher model path support.

The interactions implementation has been completely replaced with an unchanged public API surface. Computer Use API fields are now exposed, Gemma 4 local tokenizer support is added, and live music API keys are kept out of websocket URLs.

Agent Platform MCP support is now available in async generate_content, and ReinforcementTuning is supported in the GenAI SDK including the ValidateReward API method. Added transcription language code configuration and TranslationConfig for live translation, plus a fix to include all fields of a single tool.

Added computer use field support for Vertex and reinforcement tuning support in the GenAI SDK. Interaction API now accepts text/csv as a supported document MIME type and enables BigQuery tool in Deep Research config.