Google Gemini API

Accepts text, images, audio, video, and code in a single `contents` array with `parts` structure, processing all modalities through a shared transformer architecture. The API normalizes heterogeneous inputs into a unified token representation before passing to the model, enabling seamless cross-modal reasoning without separate preprocessing pipelines. Supports inline media (base64-encoded) and URI-based references for cloud-hosted assets.

Maintains a 1M+ token context window per request, allowing developers to include entire codebases, long documents, or multi-turn conversation histories in a single prompt. Context caching (paid feature) stores frequently-reused context (e.g., system prompts, reference documents) server-side for 5 minutes, charging $0.20 per 1M cached tokens plus $4.50/1M tokens/hour storage, reducing redundant token processing by up to 90% for repeated queries against the same context.

Provides free API access to limited Gemini models (specific models unknown) with unspecified token quotas and rate limits. Free tier requires no billing account initially but content is used to improve Google products (opt-out requires paid tier activation). Grounding (Google Search/Maps) includes 5,000 free queries/month shared across all Gemini 3 models before $14/1,000 query charges apply.

Web-based IDE (https://aistudio.google.com) for interactive prompt development, model testing, and API exploration without writing code. Supports multimodal input (text, images, code), real-time model response preview, prompt history, and one-click API code generation (Python, JavaScript, Go, Java, C#, REST). Enables non-technical users to prototype and technical users to iterate on prompts before integrating into applications.

Lightweight Gemini Nano model optimized for on-device inference on Android and Chrome browsers, enabling local LLM execution without cloud API calls. Reduces latency (sub-100ms inference), eliminates network dependency, and preserves privacy by keeping data on-device. Suitable for real-time applications (autocomplete, live translation) and offline-first use cases.

Exposes all API functionality via REST endpoints, enabling integration without SDKs using any HTTP client (curl, fetch, requests, etc.). Primary endpoint is `POST https://generativelanguage.googleapis.com/v1beta/models/{model}:generateContent`, accepting JSON request bodies with `contents`, `tools`, `responseSchema`, and other parameters. Responses are JSON objects with `candidates` array containing generated content. Authentication uses API key in `x-goog-api-key` header or query parameter.

Enables structured tool invocation through a schema-based function registry where developers define tool signatures as JSON schemas; the model generates structured function calls matching the schema, which SDKs automatically parse and return as callable objects. Supports native bindings for OpenAI, Anthropic, and Ollama function-calling APIs, allowing drop-in replacement of provider-specific implementations without application-level refactoring.

Executes Python code generated by the model in a sandboxed runtime environment and automatically injects execution results back into the conversation context. The model can iteratively refine code based on execution output (errors, print statements, variable values) without requiring external code execution infrastructure. Supports standard Python libraries and provides access to file I/O and system operations within sandbox constraints.

Augments model responses with real-time information from Google Search or Google Maps APIs, allowing the model to cite current events, prices, locations, or other time-sensitive data. Grounding is invoked per-request via API parameters; Google's infrastructure handles search query formulation, result ranking, and citation attribution. Paid feature: 5,000 queries/month free (shared across Gemini 3 models), then $14 per 1,000 queries.

Constrains model output to match a developer-defined JSON schema, ensuring responses conform to expected structure (e.g., specific fields, data types, required properties). The API validates generated output against the schema before returning to the client; if validation fails, the model regenerates output until it matches. Supports nested objects, arrays, enums, and type constraints (string, number, boolean, etc.).

Processes multiple requests asynchronously in a single batch job, reducing per-token costs by 50% compared to real-time API calls. Developers submit batch files (JSONL format with multiple requests), poll for job completion, and retrieve results. Batch processing trades latency (hours to days) for cost savings, suitable for non-time-sensitive workloads like daily report generation, bulk data processing, or overnight analysis.

Enables the model to decompose complex tasks into subtasks, plan execution sequences, and iteratively refine plans based on intermediate results. The model can reason about task dependencies, estimate effort, and adjust strategy mid-execution. Supports multi-step workflows where the model decides which tools to invoke, in what order, and how to handle failures or unexpected results.

Preview feature enabling the model to conduct multi-step research investigations with collaborative planning (model and user iterate on research direction), visualization generation (charts, diagrams), and MCP (Model Context Protocol) support for tool integration. The model can plan research sequences, gather information from multiple sources, synthesize findings, and generate visual summaries without requiring external research tools.

Provides native SDKs for Python, JavaScript/TypeScript, Go, Java, and C# with idiomatic API design for each language (e.g., async/await in JavaScript, context-based patterns in Go, exception handling in Java). SDKs handle authentication, request serialization, response parsing, and error handling, reducing boilerplate code. REST API available for unsupported languages or custom integrations.