Google just unveiled Gemini 3.5 Flash at Google I/O 2026, and it might be the most important AI model release of the year so far. The new model outperforms Gemini 3.1 Pro on nearly every benchmark while running four times faster than competing frontier models. But there is a catch: it costs three times more than its predecessor.
Announced on May 19, 2026, Gemini 3.5 Flash is now the default AI model across the Gemini app, AI Mode in Google Search, and the newly launched Gemini Spark personal agent. It represents a strategic pivot for Google, shifting from conversational AI to what the company calls “agentic AI,” where models do not just answer questions but autonomously plan, build, and execute real work.
Here is everything you need to know about Gemini 3.5 Flash, including its benchmarks, pricing, how it compares to Claude and GPT, and why Google is betting its AI future on agents rather than chatbots.
What Is Gemini 3.5 Flash?
Gemini 3.5 Flash is the first model in Google’s new Gemini 3.5 family. It is a multimodal AI model that accepts text, images, audio, and video input and generates text output. It features a 1 million token context window and supports up to 65,536 output tokens.
Key specs at a glance:
Release date: May 19, 2026 (generally available)
API model ID: gemini-3.5-flash
Context window: 1,048,576 input tokens
Max output: 65,536 tokens
Modalities: Text, image, audio, video input; text output
Thinking levels: Minimal, low, medium (default), high
Pricing: $1.50 per 1M input tokens / $9.00 per 1M output tokens
Cached input discount: 90%
According to Google’s official blog, the model is built for agentic workflows and coding tasks, with a focus on handling complex, multi-step projects that produce tangible results. The company also confirmed that Gemini 3.5 Pro is being used internally and will roll out next month.
Gemini 3.5 Flash Benchmarks: How It Stacks Up Against the Competition
The benchmark numbers are where things get interesting. Gemini 3.5 Flash does not just improve on its predecessor; it surpasses Google’s own flagship Gemini 3.1 Pro model on multiple fronts.
Coding and agentic benchmarks:
Terminal-Bench 2.1: 76.2%
GDPval-AA (real-world agentic tasks): 1,656 Elo
MCP Atlas (scaled tool use): 83.6%
CharXiv Reasoning (multimodal): 84.2%
Artificial Analysis Intelligence Index comparison:
ModelIntelligence IndexSpeed (tok/s)Input $/1MOutput $/1MGPT-5.5 (xhigh)60N/AN/AN/AClaude Opus 4.7 (max)57N/AN/AN/AGemini 3.5 Flash (high)55284$1.50$9.00Grok 4.3 (high)53N/AN/AN/AClaude Sonnet 4.6 (max)52N/AN/AN/A
According to Artificial Analysis, Gemini 3.5 Flash scores 55 on the Intelligence Index, placing it ahead of both Grok 4.3 and Claude Sonnet 4.6. The biggest gains came from agentic evaluations and a 31-point reduction in hallucination rate compared to Gemini 3 Flash.
The result also adds another layer to the ongoing Claude vs ChatGPT debate, especially as both OpenAI and Anthropic continue pushing higher-end reasoning models.
The multimodal gap is worth highlighting. Gemini 3.5 Flash scored 84% on MMMU-Pro, the highest score ever recorded on that benchmark. Unlike Claude Opus 4.7, Grok 4.3, and GPT-5.5, which only support image input alongside text, Gemini 3.5 Flash natively processes images, video, speech, and text together.
That makes Gemini’s multimodal edge especially relevant in the wider AI model showdown between Grok, ChatGPT, Gemini, and DeepSeek.
At 284 output tokens per second, Google CEO Sundar Pichai stated at the I/O keynote that the model is “four times faster than other frontier models.” An optimized version reportedly pushes that figure to 12 times faster with equivalent quality.
Gemini 3.5 Flash Pricing: 3x More Expensive Than Gemini 3 Flash
Here is the part that developers and enterprises need to pay attention to. Gemini 3.5 Flash costs significantly more than its predecessor:
Gemini 3 Flash: $0.50 input / $3.00 output per 1M tokens
Gemini 3.5 Flash: $1.50 input / $9.00 output per 1M tokens
That is a 3x price increase across the board. The 90% discount on cached input tokens ($0.15 per 1M) softens the blow for repetitive workloads, but the total cost to run comprehensive evaluations has jumped 5.5x compared to Gemini 3 Flash.
For enterprises, this pricing change signals that Google no longer views Flash-tier models as budget alternatives. They are now premium products that happen to be faster than Pro models. Google frames this differently, arguing that enterprises shifting 80% of workloads to a Flash mix could save over $1 billion annually at high scale, compared to using Pro-tier models exclusively.
The tradeoff is clear: you get Pro-level intelligence at Flash-tier speed, but you pay more than the old Flash pricing. Whether that math works depends entirely on your specific workload and volume.
The Agentic AI Shift: From Chatbots to Autonomous Builders
The most significant part of the Gemini 3.5 Flash launch is not the model itself; it is what Google is building around it.
According to TechCrunch, this launch marks Google’s clearest pivot yet from building chatbots to building autonomous AI agents. During internal tests, Gemini 3.5 Flash agents reportedly built an operating system entirely from scratch, launching 93 separate sub-agents, generating 2.6 billion tokens, and completing the core framework in roughly 12 hours.
Three products announced alongside the model make the agentic strategy concrete:
Antigravity 2.0 is Google’s agent-first development platform, now available as a standalone desktop application. It includes a CLI for terminal-first developers, an SDK for custom agent behaviors, and integrations with Google AI Studio, Firebase, and Android. DeepMind’s chief technologist Koray Kavukcuoglu explained that Flash 3.5 was co-developed with Antigravity to give agents a dedicated workspace where they can operate and iterate independently.
Gemini Spark is a 24/7 personal AI agent powered by Gemini 3.5 Flash. It runs on dedicated Google Cloud virtual machines, meaning it continues working even when users close their laptops or lock their phones. Spark can handle recurring tasks across Gmail, Docs, Slides, and other Workspace tools, draft follow-up messages, organize information from emails, and create project documents. Beta access begins next week for Google AI Ultra subscribers in the US at $100 per month.
Managed Agents in the Gemini API allow developers to deploy agent workflows directly through the API, with native support for the Model Context Protocol (MCP) for third-party integrations.
The competitive context matters here. Anthropic has Claude Code and Cowork, OpenAI has Codex, and now Google has Antigravity 2.0. . But Google’s advantage is integration: Spark gets native access to Gmail, Calendar, Docs, and the entire Workspace suite without additional setup, something rival agents need to configure separately.
Google I/O 2026: The Bigger Picture
Gemini 3.5 Flash was not the only major announcement at I/O 2026. Google also revealed:
Gemini Omni: A new multimodal world model that generates video from text, image, and video inputs, with conversational editing capabilities
900 million monthly active users on the Gemini app (up from 400 million last year)
1 billion+ monthly active users on AI Mode in Search
3.2 quadrillion tokens processed per month across Google’s platforms, a 7x increase year-over-year
That last number deserves emphasis. Google is processing 3.2 quadrillion tokens monthly. At that scale, even small efficiency improvements in models translate to massive cost savings, which explains why the Flash-tier model getting Pro-level intelligence is such a strategic move.
What This Means for Developers, Enterprises, and Users
For developers, near-Pro reasoning at Flash-tier speed means you no longer need to choose between quality and latency for most tasks. The higher pricing, however, means careful cost management becomes essential, especially for high-volume agentic workloads where token usage compounds across multiple agent turns.
For enterprises, Flash intelligence combined with Antigravity 2.0 and Managed Agents creates a path to automating multi-step workflows. Google reports that partner banks and fintechs are already automating multi-week processes using Flash-powered agents, though specific results have not been independently verified.
For everyday users, the impact comes through Gemini Spark and the upgraded AI Mode in Search. The promise is an AI that does not just answer questions but actively manages tasks in the background.
FAQs
Is Gemini 3.5 Flash free to use?
Gemini 3.5 Flash is free in the Gemini app and AI Mode in Google Search for general users. For API access, it costs $1.50 per million input tokens and $9.00 per million output tokens. Cached input tokens receive a 90% discount at $0.15 per million.
How does Gemini 3.5 Flash compare to Claude Sonnet 4.6?
Gemini 3.5 Flash scores 55 on the Artificial Analysis Intelligence Index, compared to Claude Sonnet 4.6’s score of 52. Gemini 3.5 Flash also supports broader multimodal input (text, image, audio, video) while Claude Sonnet 4.6 currently supports text and image input only. Speed-wise, Gemini 3.5 Flash generates over 280 tokens per second.
What is the difference between Gemini 3.5 Flash and Gemini 3.5 Pro?
Gemini 3.5 Flash is optimized for speed and agentic tasks, delivering near-Pro intelligence at lower latency. Gemini 3.5 Pro, which Google is using internally and plans to release next month, will offer deeper reasoning capabilities for tasks requiring maximum accuracy and context understanding. The two models are designed to work together in Google’s agent ecosystem.
What is Gemini Spark?
Gemini Spark is Google’s new 24/7 personal AI agent powered by Gemini 3.5 Flash. It runs on Google Cloud infrastructure and can autonomously handle tasks across Gmail, Docs, Slides, and other Workspace tools, even when your device is off. It is rolling out in beta to US Google AI Ultra subscribers.
Is Gemini 3.5 Flash better than GPT-5.5?
GPT-5.5 currently leads the Intelligence Index with a score of 60, compared to Gemini 3.5 Flash’s 55. However, Gemini 3.5 Flash is significantly faster at over 280 tokens per second and supports broader multimodal inputs including video and audio. The choice depends on whether you prioritize raw intelligence or speed and multimodal capabilities.
