Gemini 3.5 Flash

Frontier performance for agents and coding, at Flash speed.

Gemini 3.5 Flash is Google’s latest fast model in the Gemini 3 family. It pairs frontier-level reasoning with Flash-level latency and cost, and is designed for agentic workflows, iterative coding, and long-horizon multi-step tasks.

It understands text, images, audio, code, and video, and works well for both real-time interactive apps and high-volume production pipelines.

What’s new vs Gemini 3 Flash

• 42% better long-range, multi-turn performance on Armadin’s cyber benchmark, with 72% fewer tokens used.
• +19.6% on Box’s enterprise work evaluation vs Gemini 3 Flash.
• +10–20% on low-reasoning coding tasks vs the previous Flash generation.
• Coding and reasoning quality close to Gemini Pro, while keeping Flash’s speed and cost profile.

Benchmarks

Gemini 3.5 Flash leads or matches frontier models across a wide range of evaluations:

What it’s good at

• Agentic coding — terminal use, multi-step tool calls, long-running tasks.
• Multimodal reasoning — synthesize information from charts, screenshots, and documents.
• Long-context tasks — strong recall and reasoning across 128k+ token inputs.
• High-volume workflows — extraction, classification, moderation, translation at scale.
• Interactive apps — low latency makes it a good fit for assistants, in-app agents, and live tools.

Inputs

• prompt — the text prompt to send to the model.
• images — up to 10 images (each up to 7 MB).
• videos — up to 10 videos (each up to 45 minutes).
• audio — a single audio file (up to 8.4 hours).
• system_instruction — optional system prompt to steer behavior.
• thinking_level — none, low, or high. Controls how much the model reasons before answering. Higher levels improve quality on hard problems at the cost of latency.
• temperature, top_p, max_output_tokens — standard sampling controls.

Pricing

You only pay for the tokens you use. There’s no charge for cold starts or idle time.