Workhorse Model Gemini 3.6 Flash: 17% Token Reduction and 65% SWE Benchmark Gain
Google upgraded its mid-tier workhorse series with the official launch of Gemini 3.6 Flash, focusing on token efficiency and multi-step agentic execution. According to Artificial Analysis index metrics, Gemini 3.6 Flash consumes 17% fewer output tokens compared to 3.5 Flash, with API pricing set at 7.50 per 1M output tokens.
The model achieved performance gains of up to 65% on software engineering benchmarks such as DeepSWE. Official evaluation data confirms that Gemini 3.6 Flash maintains a lower output token overhead while delivering significant throughput advantages in complex code completion and agentic tool-calling tasks. Concurrently, the AI coding assistant Antigravity has fully integrated Gemini 3.6 Flash, allowing developers to directly select and leverage the model for agentic coding workflows.
High-Throughput 3.5 Flash-Lite: 350 tokens/sec Speed and Computer Use Integration
Targeting low-latency workloads like bulk document parsing and agentic web search, Google introduced Gemini 3.5 Flash-Lite as the fastest model in the 3.5 family. Output generation speed reaches 350 tokens per second, with pricing set at 2.50 per 1M output tokens, delivering a major quality jump over its 3.1 Flash-Lite predecessor.
Functionally, the model incorporates native Computer Use tools. Developers can dynamically configure "thinking levels" via API to balance latency and token cost against complex reasoning demands.
Domain-Specific 3.5 Flash Cyber: Multi-Agent Automated Vulnerability Remediation
For vertical defense applications, Google released Gemini 3.5 Flash Cyber, a specialized model fine-tuned for cybersecurity operations. Deeply integrated into the CodeMender agent platform, the model enables multi-agent collaboration to analyze security telemetry and generate verified code patches.
Currently, Gemini 3.5 Flash Cyber is accessible via a limited pilot program for government and trusted partners, providing automated code security audits at a token cost far below general-purpose frontier models.