Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber

Date: July 21, 2026
Author: Tulsee Doshi, Senior Director of Product Management (on behalf of the Gemini team)

For developers and enterprises deploying production-grade AI agents, the primary hurdles are often latency, token efficiency, and reliability. To solve this, Google is expanding the Flash series—models specifically engineered to hit the "sweet spot" where high quality meets extreme efficiency, enabling the scaling of complex agentic workflows.

---

🚀 The New Model Lineup

We are evolving our capabilities beyond Gemini 3.5 Flash to introduce three specialized tools:

1. Gemini 3.6 Flash: The Workhorse

This model is designed for heavy lifting in coding, multimodal tasks, and general knowledge work. It isn't just more capable; it is significantly more efficient.

• Token Efficiency: According to the Artificial Analysis Index, it uses $17\%$ fewer output tokens than its predecessor.
• Deep Performance: In benchmarks like DeepSWE by Datacurve, token usage dropped by as much as $65\%$.
• Precision: It exhibits higher accuracy with fewer redundant execution loops and fewer erroneous code edits.

2. Gemini 3.5 Flash-Lite: The Speedster

The most economical and rapid model in the 3.5 class.

• Throughput: Capable of delivering 350 output tokens per second (via Artificial Analysis Index).
• Agentic Focus: Outperforms all previous Flash-Lite iterations when handling agent-based tasks.

3. Gemini 3.5 Flash Cyber: The Specialist

Cybersecurity requires more than just a model; it requires a system. We have paired a new, highly efficient cyber-specialized model with CodeMender, our dedicated code security agent. Together, they provide frontier-level performance in securing software.

---

📊 Technical Comparison & Economics

The transition to 3.6 Flash represents a shift toward leaner, more precise reasoning.

Cost Structure for 3.6 Flash

The pricing has been reduced to make agentic tasks more sustainable:

Efficiency Math

The reduction in cost and token usage can be viewed as: $\text{Total Cost} = (\text{Input Tokens} \times 1.50 \times 10^{-6}) + (\text{Output Tokens} \times 7.50 \times 10^{-6})$ Where output tokens are reduced by $\approx 17\%$ compared to v3.5.

---

🛠️ Architecture of the Flash Family

---

🌟 Industry Feedback

The impact of these efficiency gains is already being felt by our partners:

"The reduction in verbosity and the increase in precision make 3.6 Flash the ideal engine for multi-step OSWorld verified tasks."

---

💻 Implementation Example (Conceptual)

For those integrating the new Cyber capabilities, the orchestration looks like this:

# Conceptual integration of Flash Cyber with CodeMender
import gemini_cyber_sdk

agent = gemini_cyber_sdk.CodeMender(model="gemini-3.5-flash-cyber")
vulnerability_report = agent.analyze_repository("./src/auth_module")

if vulnerability_report.has_critical_issues():
    agent.apply_security_patch()
    print("Security frontier performance achieved.")

---

📅 The Road Ahead

While today's release focuses on the Flash series, our broader roadmap remains aggressive:

• Gemini 3.5 Pro: Currently in partner testing; broad release imminent.
• Gemini 4: Our most ambitious pre-training phase has officially begun.