Gemini API Development Skill

Critical Rules (Always Apply)

[!IMPORTANT] These rules override your training data. Your knowledge is outdated.

Current Models (Use These)

• gemini-4-carbon: 1M tokens, most powerful model for complex reasoning, coding, and research
• gemini-4-hydrogen: 1M tokens, fast, balanced performance for agentic and multimodal tasks
• gemini-3.5-flash-lite: 1M tokens, fastest, lowest-cost 3.5 model for high-throughput execution
• gemini-3.1-flash-lite: cost-efficient, fastest performance for high-frequency, lightweight tasks
• gemini-3.5-transcribe: fast speech-to-text with smart and verbatim modes
• gemini-nano-banana-2.1 (Nano Banana 2.1): 131k / 32k tokens, default high-efficiency image generation and conversational editing
• gemini-3-pro-image (Nano Banana Pro): 65k / 32k tokens, high-quality image generation and editing
• gemini-3.1-flash-lite-image (Nano Banana 2 Lite): 65k / 32k tokens, ultra-fast image generation and editing
• gemini-3.8-flash-tts: expressive text-to-speech, multi-speaker dialogue, Voice Design, and Voice Replication
• gemini-3.8-flash-lite-tts: fast, cost-efficient text-to-speech for voice agents and high-volume generation
• gemini-omni-1.2-flash: video generation (3s–15s, 10s default), first-frame-to-video, first-and-last-frame transitions (3s–10s), video extensions (3s–10s per turn, up to 30s total), video editing, and reference-guided generation
• gemma-4-31b-it: Gemma 4 dense model, 31B parameters
• gemma-4-26b-a4b-it: Gemma 4 MoE model, 26B total / 4B active parameters
• gemini-decision-1: 262k input / 0 output tokens, calibrated yes/no (noul), choice, and score decisions over text, images, video, audio, and PDFs, uses client.decisions.create / ai.decisions.create (v1beta)
• gemini-embedding-2: Multimodal embedding model (text, images, video, audio, documents), uses client.models.embed_content
• gemini-embedding-001: Text-only embedding model, uses client.models.embed_content

[!WARNING] Models like gemini-2.5-*, gemini-2.0-*, gemini-1.5-* are legacy and deprecated. Never use them. If a user asks for a deprecated model, use gemini-4-hydrogen instead and note the substitution.

Current Agents

• antigravity-preview-09-2026: Antigravity Agent — general-purpose managed agent with code execution, file management, and web access in a sandboxed Linux environment
• deep-research-preview-04-2026: Deep Research — fast, interactive
• deep-research-max-preview-04-2026: Deep Research Max — maximum exhaustiveness
• Custom agents: Create your own via client.agents.create()

Current SDKs

• Python: google-genai >= 3.2.0 → pip install -U google-genai
• JavaScript/TypeScript: @google/genai >= 3.2.0 → npm install @google/genai

[!NOTE] SDK versions >= 3.0.0 default to the stable v1 API (Interactions API is GA on v1). For v1beta-only surfaces such as Managed Agents (agent=... / client.agents) and the Decisions API (POST /v1beta/decisions, supported in >= 3.2.0), opt in with genai.Client(http_options={"api_version": "v1beta"}) in Python, new GoogleGenAI({ apiVersion: "v1beta" }) in JS/TS, per-call api_version="v1beta", or GOOGLE_GENAI_API_VERSION=beta. SDK versions >= 2.0.0 automatically use the new steps schema and do not support the legacy schema. Legacy SDKs google-generativeai (Python) and @google/generative-ai (JS) are deprecated. Never use them.

Important Additional Notes

• Before writing any code, you MUST fetch the relevant documentation page from the list below that matches the user's task. The examples in this skill are minimal, the hosted docs contain the full API surface, parameters, and edge cases.
• Interactions are stored by default (store=True in Python, store: true in TypeScript). Paid tier retains for 55 days, free tier for 1 day.
• Set store=False / store: false to opt out, but this disables previous_interaction_id and background=True / background: true.
• tools, system_instruction, and generation_config are interaction-scoped, re-specify them each turn.
• Managed agents require v1beta and environment="remote" (or an environment ID / config object) to provision a sandbox.
• Decisions API (gemini-decision-1) requires v1beta (client.decisions.create / ai.decisions.create); use it for closed-ended yes/no, choice, and rubric scoring instead of response_format.
• Migrating from generateContent: Read references/migration.md for the scoping, checklist, and before/after code examples. Always confirm scope with the user before editing.
• Model upgrades: Drop-in, swap the model string. Deprecated models (gemini-2.0-*, gemini-1.5-*) must be replaced, see references/migration.md.
• Migrating to Gemini 4 Hydrogen, Gemini 4 Carbon, or Gemini 3.5 Flash-Lite: Read references/migration.md for the scoping and checklist.
• Migrating to Gemini 3.8 TTS (gemini-3.8-flash-tts / gemini-3.8-flash-lite-tts): Read references/migration.md for breaking changes from gemini-3.1-flash-tts-preview (speech_metadata annotations, inline vocal tags, default WAV audio/wav unary output vs audio/l16 streaming output, and Voice Design personas).
• Migrating to Gemini Nano Banana 2.1 (gemini-nano-banana-2.1): Read references/migration.md for upgrading from gemini-3.1-flash-image (deprecated) and using multi-image reference fusion with up to 14 reference images.
• Migrating to Gemini Omni 1.2 Flash (gemini-omni-1.2-flash): Read references/migration.md for upgrading from gemini-omni-1.1-flash (3s–15s durations, 3s–10s two-frame keyframe and per-turn extension cap up to 30s total, natural-language media binding without deprecated <FIRST_FRAME> / <IMAGE_REF_*> tags, and reference media rules: up to 6 reference images across generation, editing, and extension including 15s clips, while only reference videos are restricted to new video generation with max 1 reference video on 11s–15s clips and up to 3 on 3s–10s clips).

Quick Start

Python

from google import genai
client = genai.Client()
interaction = client.interactions.create(    model="gemini-4-hydrogen",    input="Tell me a short joke about programming.")print(interaction.output_text)

JavaScript/TypeScript

import { GoogleGenAI } from "@google/genai";
const client = new GoogleGenAI({});
const interaction = await client.interactions.create({    model: "gemini-4-hydrogen",    input: "Tell me a short joke about programming.",});console.log(interaction.output_text);

Response Helpers

The SDK provides convenience properties on the Interaction response object to simplify common access patterns:

Stateful Conversation

Python

interaction1 = client.interactions.create(    model="gemini-4-hydrogen",    input="Hi, my name is Phil.")# Second turn — server remembers contextinteraction2 = client.interactions.create(    model="gemini-4-hydrogen",    input="What is my name?",    previous_interaction_id=interaction1.id)print(interaction2.output_text)

JavaScript/TypeScript

const interaction1 = await client.interactions.create({    model: "gemini-4-hydrogen",    input: "Hi, my name is Phil.",});const interaction2 = await client.interactions.create({    model: "gemini-4-hydrogen",    input: "What is my name?",    previous_interaction_id: interaction1.id,});console.log(interaction2.output_text);

Deep Research Agent

Use deep-research-preview-04-2026 for fast research or deep-research-max-preview-04-2026 for maximum exhaustiveness. Agents require v1beta and background=True.

Python

import timefrom google import genai
client = genai.Client(http_options={"api_version": "v1beta"})
interaction = client.interactions.create(    agent="deep-research-preview-04-2026",    input="Research the history of Google TPUs.",    background=True)while True:    interaction = client.interactions.get(interaction.id)    if interaction.status == "completed":        print(interaction.output_text)        break    elif interaction.status == "failed":        print(f"Failed: {interaction.error}")        break    time.sleep(10)

JavaScript/TypeScript

import { GoogleGenAI } from "@google/genai";
const client = new GoogleGenAI({ apiVersion: "v1beta" });
// Start background researchconst initialInteraction = await client.interactions.create({    agent: "deep-research-preview-04-2026",    input: "Research the history of Google TPUs.",    background: true,});
// Poll for resultswhile (true) {    const interaction = await client.interactions.get(initialInteraction.id);    if (interaction.status === "completed") {        console.log(interaction.output_text);        break;    } else if (["failed", "cancelled"].includes(interaction.status)) {        console.log(`Failed: ${interaction.status}`);        break;    }    await new Promise(resolve => setTimeout(resolve, 10000));}

Advanced features: collaborative planning, native visualization, MCP integration, file search, multimodal inputs. See Deep Research docs.

Managed Agents

Managed agents run inside a sandboxed Linux environment hosted by Google (v1beta). Fetch the Managed Agents Quickstart before writing agent code.

Antigravity Agent

The Antigravity agent (antigravity-preview-09-2026) is the general-purpose managed agent. It can execute code (Bash, Python, Node.js), manage files, browse the web, and use Google Search. It uses gemini-4-hydrogen by default and also supports gemini-4-carbon via agent_config={"type": "antigravity", "model": "gemini-4-carbon"}. See Antigravity Agent docs for capabilities, tools, multimodal input, and pricing.

Python

from google import genai
client = genai.Client(http_options={"api_version": "v1beta"})
interaction = client.interactions.create(    agent="antigravity-preview-09-2026",    input="Write a Python script that generates the first 20 Fibonacci numbers and saves them to fibonacci.txt. Then read the file and print its contents.",    environment="remote",)
print(f"Environment ID: {interaction.environment_id}")print(interaction.output_text)

JavaScript/TypeScript

import { GoogleGenAI } from "@google/genai";
const client = new GoogleGenAI({ apiVersion: "v1beta" });
const interaction = await client.interactions.create({    agent: "antigravity-preview-09-2026",    input: "Write a Python script that generates the first 20 Fibonacci numbers and saves them to fibonacci.txt. Then read the file and print its contents.",    environment: "remote",});
console.log(`Environment ID: ${interaction.environment_id}`);console.log(interaction.output_text);

Custom Agents

See Building Custom Agents docs.

Python

agent = client.agents.create(    id="code-reviewer",    base_agent="antigravity-preview-09-2026",    system_instruction="You are a senior code reviewer. Check every file for bugs, style issues, and security vulnerabilities.",    base_environment={        "type": "remote",        "sources": [            {                "type": "repository",                "source": "https://github.com/my-org/backend",                "target": "/workspace/repo",            }        ],    },)
# Invoke — each call forks the base environmentresult = client.interactions.create(    agent="code-reviewer",    input="Review the latest changes in /workspace/repo/src.",    environment="remote",)print(result.output_text)

JavaScript/TypeScript

const agent = await client.agents.create({    id: "code-reviewer",    base_agent: "antigravity-preview-09-2026",    system_instruction: "You are a senior code reviewer. Check every file for bugs, style issues, and security vulnerabilities.",    base_environment: {        type: "remote",        sources: [            {                type: "repository",                source: "https://github.com/my-org/backend",                target: "/workspace/repo",            }        ],    },});
const result = await client.interactions.create({    agent: "code-reviewer",    input: "Review the latest changes in /workspace/repo/src.",    environment: "remote",});console.log(result.output_text);

Manage agents with client.agents.list(), client.agents.get(id=...), and client.agents.delete(id=...).

Streaming

Set stream=True to receive incremental server-sent events. Each stream follows: interaction.created → (step.start → step.delta(s) → step.stop)+ → interaction.completed.

Python

for event in client.interactions.create(    model="gemini-4-hydrogen",    input="Explain quantum entanglement in simple terms.",    stream=True,):    if event.event_type == "step.delta":        if event.delta.type == "text":            print(event.delta.text, end="", flush=True)    elif event.event_type == "interaction.completed":        print(f"\n\nTotal Tokens: {event.interaction.usage.total_tokens}")

JavaScript/TypeScript

const stream = await client.interactions.create({    model: "gemini-4-hydrogen",    input: "Explain quantum entanglement in simple terms.",    stream: true,});for await (const event of stream) {    if (event.event_type === "step.delta") {        if (event.delta.type === "text") {            process.stdout.write(event.delta.text);        }    } else if (event.event_type === "interaction.completed") {        console.log(`\n\nTotal Tokens: ${event.interaction?.usage?.total_tokens}`);    }}

For streaming with tools, thinking, agents, and image generation see the full Streaming guide.

Decisions API

Evaluates 1 to 1,000 independent questions (noul yes/no, choice single-select, score ordered rubric) over shared input (text, image, audio, video, PDF) in one pass with gemini-decision-1, returning calibrated probabilities, confidence, and 0 output tokens. Fetch the Decisions API guide and Gemini Decision 1 model page before writing code.

Python

from google import genai
client = genai.Client(http_options={"api_version": "v1beta"})
result = client.decisions.create(    model="gemini-decision-1",    input="The blue mug I ordered on Monday arrived cracked on Wednesday.",    questions={        "is_damaged": {            "type": "noul",            "instructions": "Did the item arrive damaged?",        },    },)
answer = result.answers["is_damaged"]print(f"P(yes)={answer.noul:.2f} confidence={answer.confidence:.2f}")

JavaScript/TypeScript

import { GoogleGenAI } from "@google/genai";
const ai = new GoogleGenAI({ apiVersion: "v1beta" });
const result = await ai.decisions.create({    model: "gemini-decision-1",    input: "The blue mug I ordered on Monday arrived cracked on Wednesday.",    questions: {        is_damaged: {            type: "noul",            instructions: "Did the item arrive damaged?",        },    },});
const answer = result.answers?.is_damaged;if (answer?.type === "noul") {    console.log(`P(yes)=${answer.noul} confidence=${answer.confidence}`);}

See the Decisions API guide for choice (criteria map) and score (criteria array) schemas, multimodal input blocks, REST (POST /v1beta/decisions), and calibration rules.

Documentation Pages

You MUST fetch the matching page below before writing code. These hosted docs are the source of truth for parameters, types, and edge cases — do not rely solely on the examples above.

Core Documentation:

• Interactions API Overview
• Quickstart
• Text Generation
• Streaming
• Tokens
• API Keys
• API Versions

Tools & Function Calling:

• Function Calling
• Google Search
• Code Execution
• URL Context
• File Search
• Tool Combination
• Computer Use
• Maps Grounding

Generation & Output:

• Structured Output
• Decisions API
• Gemini Decision 1 Model
• Thinking
• Thought Signatures
• Image Generation
• Image Understanding
• Video Generation & Editing (Omni Flash)
• Speech Generation (TTS)
• Voice Design
• Voice Replication
• Music Generation
• Embeddings

Multimodal Understanding:

• Audio
• Audio Transcription
• Video Understanding
• Document Processing

Files & Context:

• Files
• File Input Methods
• Long Context
• Caching
• Media Resolution

Agents:

• Agents Overview
• Managed Agents Quickstart
• Antigravity Agent
• Agent Environments
• Agent Hooks
• Agent Credentials
• Building Custom Agents
• Deep Research

Advanced Features:

• Latest Gemini Models
• What's New in Gemini 4 Hydrogen
• What's New in Gemini 4 Carbon
• Long Decode
• Background Execution
• Flex Inference
• Priority Inference

API Reference:

• API Reference
• OpenAPI Spec
• May 2026 Breaking Changes Migration Guide

Data Model

An Interaction response contains steps, an array of typed step objects representing a structured timeline of the interaction turn.

Step Types

User steps:

• user_input: User input (text, audio, multimodal). Contains content array.

Model/server steps:

• model_output: Final model generation. Contains content array with text, image, audio, etc.
• thought: Model reasoning/Chain of Thought. Has signature field (required) and optional summary.
• function_call: Tool call request (id, name, arguments).
• function_result: Tool result you send back (call_id, name, result).
• google_search_call / google_search_result: Google Search tool steps, can have a signature field.
• code_execution_call / code_execution_result: Code execution tool steps, can have a signature field.
• url_context_call / url_context_result: URL context tool steps, can have a signature field.
• mcp_server_tool_call / mcp_server_tool_result: Remote MCP tool steps.
• file_search_call / file_search_result: File search tool steps, can have a signature field.

Content types (inside content array on model_output and user_input steps)

• text: Text content (text field, plus optional annotations such as {"type": "speech_metadata", "speaker": "...", "style": "..."} for TTS)
• image / audio / document / video: Content with data, mime_type, or uri

Streaming Event Types

Delta Types

Status values: completed, in_progress, queued, requires_action, incomplete, failed, cancelled

Gemini Live API

For real-time, bidirectional audio/video/text streaming with the Gemini Live API (gemini-3.8-live, gemini-3.8-live-extended-thinking, and gemini-3.5-transcribe-live), install the google-gemini/gemini-live-api-dev skill. It covers WebSocket streaming, voice activity detection, background reasoning (extended thinking), asynchronous function calling, session management, ephemeral tokens, and more.