Google Gemini API

Accepts text, images, audio, video, and code in a single request via a unified parts-based content model, processing them through a shared transformer architecture that maintains semantic relationships across modalities. The API uses a standardized contents/parts JSON structure where each part can be a different media type, enabling seamless cross-modal reasoning without separate preprocessing pipelines or format conversion.

Supports prompts and responses up to 1 million tokens through a transformer architecture optimized for long-context attention. Pricing is tiered at the 200K token boundary, with input costs doubling and output costs increasing 50% for contexts exceeding 200K tokens, incentivizing efficient context management while enabling retrieval-augmented generation with full document sets.

Enables the model to decompose complex tasks into multiple steps, decide which tools to call at each step, and execute a plan across multiple API calls. The model reasons about task decomposition, tool selection, and execution order, with the client orchestrating the execution loop by feeding tool results back to the model for the next step.

Supports generation and understanding in 24+ languages including English, German, Spanish, French, Indonesian, Italian, Polish, Portuguese, Turkish, Russian, Hebrew, Arabic, Persian, Hindi, Bengali, Thai, Simplified Chinese, Traditional Chinese, Japanese, Korean, and others. The model handles language detection, translation, and code-switching without explicit language specification, enabling multilingual applications.

Provides Gemini Nano, a lightweight model optimized for on-device execution on Android and Chrome platforms, enabling low-latency, privacy-preserving inference without cloud API calls. The model runs directly on the user's device, eliminating network latency and keeping data local, though with reduced capabilities compared to cloud Gemini models.

Provides free API access via Google AI Studio with limited model availability (only 'some' models), free input and output tokens (quota limits unknown), and content used for product improvement. The free tier enables prototyping and low-volume use without payment, though with restrictions on model selection, token quotas, and data privacy.

Provides a Priority tier with 3.6x standard pricing that guarantees lower latency and higher throughput for time-sensitive applications. Requests are processed with higher priority in the queue, reducing wait times and enabling consistent sub-second response times for production applications that require predictable performance.

Provides an Enterprise tier with provisioned throughput (custom capacity reserved for the customer), volume-based discounts (custom pricing based on usage), and dedicated support. Enterprises can negotiate custom SLAs, guaranteed capacity, and discounted per-token rates based on volume commitments.

Enables the model to invoke external functions by declaring tool schemas (function signatures, parameters, descriptions) in the request, with the API returning structured tool calls that clients execute and feed back as tool results. The implementation uses a schema-based registry pattern where tools are defined declaratively, allowing the model to reason about which tools to call and in what order without hardcoded tool logic.

Constrains model outputs to conform to a provided JSON schema, ensuring responses are valid, parseable structured data suitable for downstream processing. The model generates text that adheres to the schema constraints, with the API validating output before returning it to the client, eliminating the need for post-processing parsing or validation.

Integrates real-time Google Search results into the generation process, allowing the model to cite current information and ground responses in verifiable sources. The API queries Google Search, retrieves relevant results, and incorporates them into the context before generation, enabling responses about recent events, current prices, or other time-sensitive information that would be outdated in the model's training data.

Integrates Google Maps data (locations, directions, business information, reviews) into the generation process, allowing the model to provide location-aware responses with current business hours, directions, or local information. Similar to Search grounding, the API queries Maps, retrieves relevant location data, and incorporates it into context before generation.

Caches large prompt contexts (system instructions, documents, code, etc.) on Google's servers, allowing subsequent requests with the same context to reuse the cached version instead of reprocessing. The API charges a one-time cache write cost ($0.20-0.40/1M tokens depending on context size) plus hourly storage costs ($4.50/1M/hour), with subsequent requests paying only for new input tokens, reducing latency and cost for applications with repeated contexts.

Accepts asynchronous batch requests via a separate Batch API endpoint, processing them at lower priority with 50% cost reduction compared to standard on-demand pricing. Clients submit batches of requests, poll for completion status, and retrieve results asynchronously, enabling cost-effective processing of non-time-sensitive workloads at half the per-token cost.

Enables the model to perform extended reasoning before generating a response by allocating 'thinking tokens' that are used for internal reasoning steps not shown to the user. The model spends thinking tokens on complex reasoning, planning, and verification before producing the final output, improving accuracy on difficult problems at the cost of additional output tokens (thinking tokens are charged at the same rate as regular output tokens).

Enables the model to write and execute code (Python, JavaScript, etc.) within the API request, with the execution environment returning results back to the model for verification or iteration. The model can generate code, execute it, see the results, and refine the code based on execution output, enabling more reliable code generation and problem-solving.