Google Unveils Next-Gen AI Models: Gemini 3.6 Flash, 3.5 Flash-Lite, and Cyber Defender 3.5 Flash Cyber

 

Organizations and software engineers deploying autonomous AI agents in production face a critical balance of priorities: maximizing token efficiency, minimizing processing latency, and maintaining reliable performance across multi-step execution paths. Google’s Flash model architecture was engineered specifically to meet this exact sweet spot, delivering an optimal combination of speed, cost efficiency, and intelligence needed to scale agentic systems. Google has announced the next evolutionary milestone in its agentic AI ecosystem by launching three specialized models: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and the security-focused Gemini 3.5 Flash Cyber. Alongside these releases, Google shared updates regarding its broader model pipeline: Gemini 3.5 Pro is currently undergoing early access partner testing with general availability anticipated soon. Simultaneously, Google DeepMind has initiated pre-training for its next-generation flagship model, Gemini 4.

---

1. Gemini 3.6 Flash: Token Efficiency and Advanced Agentic Coding

Serving as the core workhorse model of the Flash lineup, Gemini 3.6 Flash incorporates developer feedback from Gemini 3.5 Flash. It delivers significant gains in software engineering capabilities, complex reasoning, and multimodal comprehension while meaningfully reducing token consumption.

Token Economics and Efficiency Gains

• 17% Reduction in Output Tokens: According to independent measurements from the Artificial Analysis Index, Gemini 3.6 Flash consumes 17% fewer output tokens compared to 3.5 Flash while generating superior response quality.
• Up to 65% Token Savings in Coding Benchmarks: In long-horizon agentic software engineering environments such as Datacurve’s DeepSWE benchmark, token usage drops by up to 65% as a result of streamlined reasoning steps and reduced redundant tool calls.
• Cost-Effective Pricing Structure:
  • Input Tokens: $1.50 per 1 Million tokens.
  • Output Tokens: $7.50 per 1 Million tokens.

Metric / Benchmark	Gemini 3.5 Flash	Gemini 3.6 Flash	Key Improvement Details
DeepSWE (Agentic Software Engineering)	37.0%	49.0%	Higher precision, fewer unwanted code modifications and execution loops
MLE Bench (ML Research Engineering)	49.7%	63.9%	Substantial performance gains in autonomous ML research workflows
OSWorld-Verified (Computer Use)	78.4%	83.0%	Enhanced GUI navigation and autonomous desktop interaction
GDPval-AA v2 (Knowledge Work)	1349	1421	Superior document parsing, financial chart analysis, and report generation

Native Computer Use and Enterprise Integration

The Computer Use feature is now accessible as a native, client-side tool across the Gemini API and Gemini Enterprise. Early adoptersu2014including legal technology provider Harvey and enterprise research platform Hebbiau2014report significant performance improvements in complex multimodal tasks such as unstructured document parsing, financial statement analysis, and automated report synthesis. Watch Video: Token Efficiency in OSWorld (3.6 Flash vs. 3.5 Flash)

Watch Video: Financial Data & Transcript Analysis (Managed Agents on AIS)

Watch Video: Multi-Agent Code Migration (AGY Environment)

Watch Video: 3D Texture Extractor Tool (Gemini App Canvas)

Performance Evaluations & Customer Perspectives

Figure 1. Token efficiency and cost per task metrics for Gemini 3.6 Flash on the Artificial Analysis Index.

Figure 2. Benchmark quality comparisons between Gemini 3.6 Flash and 3.5 Flash across DeepSWE, MLE Bench, OSWorld, and GDPval.

Enhanced Frontier Safety Alignment

Gemini 3.6 Flash ships with updated Frontier Safety safeguards designed to mitigate Chemical, Biological, Radiological, and Nuclear (CBRN) threats as well as offensive cyber misuses. The model demonstrates substantially stronger resistance to adversarial jailbreak attempts while being specifically tuned to minimize false refusals for benign, legitimate user prompts.

---

2. Gemini 3.5 Flash-Lite: Low-Latency Model for High-Throughput Systems

For applications demanding rapid execution speed and high throughputu2014such as real-time agentic search, document indexing, and large-scale data extractionu2014Google introduced Gemini 3.5 Flash-Lite.

Generation Speed, Pricing, and Adjustable Thinking Levels

• Throughput Speed: 350 output tokens per second (measured by Artificial Analysis), making it the fastest model in the 3.5 family.
• Cost Economics:
  • Input Tokens: $0.30 per 1 Million tokens.
  • Output Tokens: $2.50 per 1 Million tokens.
• Configurable Reasoning Intensity: Developers can tune the model’s thinking budget. Minimal and low thinking levels prioritize immediate response times for high-volume background operations, whereas higher thinking levels enable deep multi-step subagent reasoning.

Benchmark / Evaluation	Gemini 3.1 Flash-Lite	Gemini 3.5 Flash-Lite	Comparison to Baseline Models
Terminal-Bench 2.1 (CLI Agentics)	31.0%	54.0%	Major improvement in command-line execution and shell tool use
GDM-MRCR v2 (Long Context)	60.1%	72.2%	Enhanced long-context information retrieval and synthesis
GDPval-AA v2 (Knowledge Work)	642	1140	Nearly double the intelligence score of prior generation Lite models
SWE-Bench Pro (Software Engineering)	u2014	54.2%	Outperforms Gemini 3 Flash (49.6%)
OSWorld-Verified (GUI Agentics)	u2014	74.0%	Outperforms Gemini 3 Flash (65.1%)

Watch Video: Latency Comparison (3.5 Flash-Lite vs. 3.5 Flash)

Watch Video: E-Commerce Product Feature Extraction at Scale

Watch Video: Receipt Scanning & Multimodal Translation

Watch Video: Agentic Puzzle Game Generation & Iteration

Evaluations & Customer Quotes

Figure 3. Performance gains of Gemini 3.5 Flash-Lite compared to 3.1 Flash-Lite.

Figure 4. Gemini 3.5 Flash-Lite outperforming Gemini 3 Flash on SWE-Bench Pro and OSWorld-Verified.

---

3. Gemini 3.5 Flash Cyber & CodeMender: Vulnerability Detection & Automated Patching

As modern AI systems become increasingly adept at identifying software security flaws faster than manual patch engineering cycles can resolve them, defensive tooling must evolve. Google is addressing this defensive gap with Gemini 3.5 Flash Cyber.

CodeMender Security Architecture

• Specialized Security Fine-Tuning: Built upon 3.5 Flash, 3.5 Flash Cyber is tailored to discover, verify, and patch security vulnerabilities at a fraction of the cost per token of larger models.
• Multi-Agent Defensive Orchestration: Operating within the CodeMender framework (engineered by Google DeepMind), multiple 3.5 Flash Cyber agents run in parallel alongside classic program analysis engines (static/dynamic analyzers, fuzzers, and SMT solvers) to produce validated code security patches.
• CyberGym Benchmark Results: The system achieves top-tier competitive results on the CyberGym cybersecurity benchmark.

Figure 5. Benchmark performance of 3.5 Flash Cyber within CodeMender on CyberGym.

Controlled Deployment Program

Due to the dual-use capabilities of cybersecurity automation, Google is deploying 3.5 Flash Cyber under strict risk governance. The model will be made available exclusively to government institutions and vetted security partners via CodeMender through a restricted-access pilot program, giving defensive teams an early advantage in patching critical vulnerabilities before exploitation.

---

4. Platform Availability and Deployment

Gemini 3.6 Flash and 3.5 Flash-Lite are available beginning July 21, 2026 across developer, enterprise, and consumer channels:

1. For Developers:
   • Accessible in the Gemini API via Google AI Studio and Android Studio.
   • Integrated into Google Antigravity and available within GitHub Copilot.
2. For Enterprise Customers:
   • Available within the Gemini Enterprise Agent Platform and the Gemini Enterprise workspace application.
3. For General Users:
   • Gemini 3.6 Flash is live in the Gemini App.
   • Gemini 3.5 Flash-Lite is rolling out inside Google Search.