QUICK VERDICT
Gemini 3.5 Flash is Google's strongest Flash model yet — faster and cheaper than Gemini 3.1 Pro while beating it on coding and agentic benchmarks. At $1.50 input / $9.00 output per million tokens, it undercuts Claude Opus 4.7 and GPT-5.5 significantly for most workloads. Best for: high-volume agentic pipelines, multi-tool coding tasks, and long-context document work. Not the best choice for pure academic reasoning benchmarks where Gemini 3.1 Pro still leads.
What Google Actually Launched at I/O 2026
Google launched Gemini 3.5 Flash at Google I/O on May 19, 2026, calling it the first model in the Gemini 3.5 family and the beginning of what CEO Sundar Pichai described as the "agentic Gemini era." It is generally available immediately — no waitlist, no preview suffix — via the Gemini API, Google AI Studio, Google Antigravity, Android Studio, and the Gemini app. It also replaced the prior default model in AI Mode in Google Search, which has now exceeded one billion monthly users.
The model supports text, image, audio, video, and PDF inputs with a 1,048,576 token context window (1M tokens) and up to 65,536 output tokens. The stable API model ID is gemini-3.5-flash — no preview suffix. A larger Gemini 3.5 Pro is in internal use at Google with a public rollout expected in June 2026.
Benchmark Results — The Numbers That Matter
| Benchmark |
Gemini 3.5 Flash |
Gemini 3.1 Pro |
What It Tests |
| Terminal-Bench 2.1 |
76.2% |
70.3% |
Coding in real terminal environments |
| GDPval-AA (Elo) |
1656 |
1314 |
Real-world agentic task performance |
| MCP Atlas |
83.6% |
78.2% |
Multi-tool coordination reliability |
| CharXiv Reasoning |
84.2% |
— |
Multimodal chart and document reasoning |
| Finance Agent v2 |
57.9% |
43.0% |
Expert financial agentic tasks |
| Humanity's Last Exam |
40.2% |
44.4% |
Pure academic reasoning depth |
| Output speed |
~280 tok/s |
~70 tok/s |
Tokens per second (Artificial Analysis) |
The pattern is consistent: 3.5 Flash is built for agentic and coding workloads and wins convincingly on those tasks. On MCP Atlas, Flash climbed from 62.0% (Gemini 3 Flash) to 83.6% — a 21-point jump that reflects how much Google prioritised tool-use reliability in this generation. The one benchmark where 3.1 Pro still leads is Humanity's Last Exam, which measures raw academic reasoning depth — useful for research tasks but less relevant for production agent pipelines.
Speed is equally notable. Artificial Analysis measured output at approximately 280 tokens per second, putting it 4x faster than comparable frontier models. For agentic pipelines where latency compounds across dozens of tool calls, that difference is practical — not just a benchmark number.
Pricing — What You Actually Pay
| Model |
Input (per 1M tokens) |
Output (per 1M tokens) |
Context |
| Gemini 3.5 Flash |
$1.50 |
$9.00 |
1M tokens |
| Gemini 3.1 Pro |
$2.00 |
$12.00 |
2M tokens |
| Gemini 3.1 Flash-Lite |
$0.25 |
$1.50 |
1M tokens |
| Claude Haiku 4.5 |
$1.00 |
$5.00 |
200K tokens |
| Claude Sonnet 4.6 |
$3.00 |
$15.00 |
1M tokens |
The headline pricing looks reasonable until you compare it to the previous Flash generation. Gemini 3 Flash Preview was $0.50 input / $3.00 output — 3.5 Flash costs three times as much. Gemini 3.1 Flash-Lite was $0.25 / $1.50 — 3.5 Flash costs six times more per token. Google is charging Pro-adjacent rates for what is technically a Flash-tier model, which has caused some grumbling among developers who relied on Flash's historically low pricing.
The justification is performance: 3.5 Flash genuinely surpasses 3.1 Pro on the benchmarks that matter for agentic work, so Google is pricing accordingly. Context caching helps significantly — cached input tokens cost $0.15 per million (a 90% reduction), and for workloads with repeated system prompts or context, caching can cut total costs by 30–50%. Non-global regions carry a slight premium at $1.65/$9.90.
Thinking Levels — How to Control Cost vs. Quality
Gemini 3.5 Flash replaces the numeric thinking budget from the previous generation with four named tiers: Minimal, Low, Medium, and High. Medium is the default — Google says it delivers strong results across most tasks while keeping speed and cost reasonable. High thinking trades latency for reasoning depth (time to first token rises significantly) and is intended for hard multi-step problems where getting the plan right matters more than raw speed.
Thinking level recommendations by task:
- Minimal — Quick classification, lightweight chat, simple lookups
- Low — Low-latency code generation, straightforward agent steps
- Medium (default) — Most production agentic pipelines, multi-turn reasoning
- High — Complex planning, security analysis, hard math — where quality matters more than latency
Note that High thinking carries a significantly elevated time to first token (Artificial Analysis measured 18.95 seconds for the High configuration), which can feel slow in chat-style applications. For agentic pipelines where intermediate thinking is invisible to the user, that tradeoff is often acceptable.
Decision Framework — Should You Switch to Gemini 3.5 Flash?
Switch to Gemini 3.5 Flash if:
- You are running agentic coding pipelines and need fast, reliable tool calling
- Your workload involves long-context documents (up to 1M tokens) at moderate cost
- You were on Gemini 3.1 Pro and want better agentic performance at 25% lower cost
- You need multimodal inputs — text, image, audio, video, PDF — in a single model
Stay on your current model if:
- You were on Gemini 3.1 Flash-Lite and cost is the primary constraint — 3.5 Flash is 6x more expensive per token
- Your use case is pure academic reasoning where Gemini 3.1 Pro's 2M context and HLE lead matter
- You need the 2M context window that only 3.1 Pro currently offers in the Gemini line
Frequently Asked Questions
What is the API model ID for Gemini 3.5 Flash?
gemini-3.5-flash — no preview suffix. Available on the Gemini API via Google AI Studio and Antigravity.
How does Gemini 3.5 Flash compare to Claude Sonnet 4.6?
Claude Sonnet 4.6 costs $3.00/$15.00 per million tokens — roughly double 3.5 Flash on both input and output. Sonnet 4.6 has a 1M context window and is stronger on instruction-following and computer use tasks. Gemini 3.5 Flash leads on multi-tool coordination (MCP Atlas: 83.6%) and is faster. For pure agentic pipeline cost-efficiency, 3.5 Flash has a meaningful price advantage. For tasks requiring Claude's instruction-following precision or computer use, Sonnet 4.6 is still the better choice.
Is Gemini 3.5 Flash free to use?
The Gemini app and AI Mode in Google Search use Gemini 3.5 Flash for free consumer users. API access is paid: $1.50/$9.00 per million tokens. Google AI Studio has a free tier with rate limits for developers testing the API.
When is Gemini 3.5 Pro releasing?
Google confirmed Gemini 3.5 Pro is already in internal use and is planned for public rollout in June 2026. No specific date has been given.
Does Gemini 3.5 Flash support the OpenAI-compatible API?
Yes. The Gemini API is OpenAI-compatible at the chat completions endpoint. You can swap the base URL and model name; the rest of your client code remains unchanged.