Gemini 3.5 Flash API: Google I/O 2026 Developer Guide

Google used Google I/O 2026 (May 19, Shoreline Amphitheatre) to ship rather than preview. Gemini 3.5 Flash became generally available the same day it was announced — no waitlist, no preview header, no regional restriction. The model is accessible via the Gemini API, AI Studio, Vertex AI, Gemini Enterprise Agent Platform, and Antigravity 2.0.

Effloow Lab verified key facts for this article from official Google developer docs, llm-stats.com, marktechpost.com, and pricepertoken.com. No live API calls were made (no GEMINI_API_KEY in this environment). The lab-run note is at data/lab-runs/gemini-3-5-flash-api-developer-guide-2026.md. All code examples use the confirmed model ID and API surface from official documentation.

What is Gemini 3.5 Flash

The naming follows a deliberate pattern in Google's model lineup. Gemini X.Y "Flash" models are the cost-efficient, high-throughput tier for production workloads — lower cost than the corresponding Pro model, optimized for latency-sensitive and agentic use cases. Gemini 3.5 Flash is positioned as the successor to Gemini 3 Flash (launched with Gemini 3 in early 2026) and the production alternative to Gemini 3.1 Pro for teams where cost and throughput matter more than maximum capability.

The headline claim: Gemini 3.5 Flash outperforms Gemini 3.1 Pro on the coding and agentic benchmark suite, at roughly 40% lower cost per token.

Model ID and API setup

Model ID: gemini-3.5-flash

The recommended SDK as of May 2026 is google-genai v2.0.0+. If you are upgrading from Gemini 3.1 Pro, the SDK version bump is required — v2.0.0 introduced breaking changes to the Interactions API.

Python (google-genai v2+)

from google import genai

client = genai.Client()

response = client.models.generate_content(
    model="gemini-3.5-flash",
    contents="What are the key differences between MCP and A2A protocols?",
)
print(response.text)

JavaScript (google-genai v2+)

import { GoogleGenAI } from '@google/genai';

const ai = new GoogleGenAI({ apiKey: process.env.GEMINI_API_KEY });

const response = await ai.models.generateContent({
  model: 'gemini-3.5-flash',
  contents: 'What are the key differences between MCP and A2A protocols?',
});
console.log(response.text);

For Vertex AI:

import vertexai
from vertexai.generative_models import GenerativeModel

vertexai.init(project="your-project-id", location="us-central1")
model = GenerativeModel("gemini-3.5-flash")
response = model.generate_content("Explain speculative decoding in 3 sentences.")
print(response.text)

Pricing

Gemini 3 Flash (the prior Flash model) was $0.50 / $3.00. Gemini 3.5 Flash is 3x more expensive, but the benchmark improvements justify the increase for agentic workloads where accuracy on tool-calling and multi-step reasoning matters. Gemini 3.1 Pro was $2.50 / $15.00 — Gemini 3.5 Flash is 40% cheaper on input and 40% cheaper on output while matching or exceeding it on the coding and agentic suite.

Benchmarks

Sources: marktechpost.com (2026-05-20), buildfastwithai.com, llm-stats.com. Comparative Gemini 3.1 Pro columns are [DATA NOT AVAILABLE] because a unified benchmark table across both models was not found in available sources at the time of writing.

Context and output limits

• Input context: 1,048,576 tokens (1M)
• Max output: 65,536 tokens
• Input modalities: text, image, audio, video
• Output modality: text

The 1M context window matches Gemini 3.1 Pro. For long-context document analysis, legal contract review, or codebase-wide tasks, the context limit is not a differentiator — the speed and cost difference is.

Agentic use: Interactions API

For stateful multi-turn agent workflows, the Interactions API is preferred over the GenerateContent API. It maintains session state server-side, reducing the need to replay long conversation histories on each turn.

from google import genai
import json

client = genai.Client()

my_tool = {
    "function_declarations": [{
        "name": "search_codebase",
        "description": "Search the codebase for a pattern",
        "parameters": {
            "type": "object",
            "properties": {
                "pattern": {"type": "string"},
                "file_glob": {"type": "string"}
            },
            "required": ["pattern"]
        }
    }]
}

# First turn
interaction = client.interactions.create(
    model="gemini-3.5-flash",
    tools=[my_tool],
    input=[{"type": "text", "text": "Find all usages of createTool in the codebase."}],
)

# Handle tool call
for step in interaction.output:
    if step.type == "function_call":
        # Execute the tool
        result = {"files": ["src/agent.ts", "src/workflow.ts"], "count": 2}

        # Continue the interaction with the tool result
        final = client.interactions.create(
            model="gemini-3.5-flash",
            previous_interaction_id=interaction.id,
            tools=[my_tool],
            input=[{
                "type": "function_result",
                "name": step.name,
                "call_id": step.id,
                "result": [{"type": "text", "text": json.dumps(result)}],
            }],
        )
        print(final.output[0].text)

Breaking change note: If you used thinking_budget in Gemini 3.1 Pro calls, the parameter is now thinking_level in the Interactions API. thinking_budget is no longer accepted and will raise a validation error.

Migration from Gemini 3.1 Pro

The primary migration steps are:

1. SDK version: upgrade to google-genai >= 2.0.0
2. Model ID: change gemini-3.1-pro to gemini-3.5-flash
3. Thinking parameter: thinking_budget → thinking_level
4. Function call responses: ensure every function_result provides matching id, name, and response fields — v2.0.0 validates all three

Token costs will decrease (40% cheaper than 3.1 Pro). Output throughput will increase (4x faster). For workloads where Gemini 3.1 Pro was the previous choice due to benchmark rankings on coding/agentic tasks, Gemini 3.5 Flash is the direct replacement — it matches or exceeds 3.1 Pro on those specific benchmarks.

Available platforms

When to use Gemini 3.5 Flash

Frequently Asked Questions

Q: Is Gemini 3.5 Flash better than Gemini 3.1 Pro for everything?

On coding and agentic benchmarks (Terminal-Bench 2.1, MCP Atlas), yes — Gemini 3.5 Flash outperforms Gemini 3.1 Pro at lower cost. For other benchmark categories (long-form reasoning, creative writing, multilingual), comparative data was not available at the time of writing. If your workload is not coding/agentic, run your own eval before switching.

Q: What is the rate limit on the free tier?

Free tier rate limits are set in Google AI Studio per project. As of launch, Gemini 3.5 Flash follows the same rate limit structure as prior Flash models. Exact per-minute and per-day limits are available at ai.google.dev/gemini-api/docs/rate-limits.

Q: Does Gemini 3.5 Flash support video input?

Yes. Like Gemini 3.1 Pro, Gemini 3.5 Flash accepts video input via the GenerateContent and Interactions APIs. Video is processed as frames; the 1M token context window determines how many frames can be processed per call.

Q: How does the 4x speed claim compare to specific frontier models?

The 4x figure comes from chatforest.com's review citing Google I/O announcements. It refers to output tokens per second compared to "comparable frontier models" — the baseline is unspecified in available sources. Treat it as directional rather than a precise benchmark. Independent throughput measurements from Artificial Analysis or LLM-Stats.com will be more reliable once they publish updated benchmarks for this model.

Q: Can I use Gemini 3.5 Flash with the Genkit framework?

Yes. Genkit 2.0 supports Gemini models through the @genkit-ai/googleai plugin. After installing the plugin, use googleAI('gemini-3.5-flash') as the model parameter in any Genkit flow.