Google Launches Three Gemini Flash Models, Including a Cybersecurity Variant That Autonomously Attacks to Verify Vulnerabilities — BigGo Finance

Google launched three Gemini Flash series AI models on July 21, US local time, pursuing higher performance and lower costs while also unveiling a cybersecurity-specific model capable of autonomously launching attacks to verify vulnerabilities—marking a new phase of active verification for AI in software security defense.

The releases include the main general-purpose model Gemini 3.6 Flash, the lightweight high-speed model Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber, fine-tuned specifically for cybersecurity vulnerability detection and remediation. All three models are designed to reduce the cost of deploying AI agents for enterprises and improve large-scale operational efficiency. However, the long-anticipated flagship model Gemini 3.5 Pro was once again absent. Google simultaneously announced it has begun pre-training work on its next-generation model, Gemini 4.

Flagship Model Cuts Costs and Boosts Efficiency, Output Token Consumption Reduced by Up to 65%

Gemini 3.6 Flash is the core product of this release. According to calculations by third-party benchmark organization Artificial Analysis Index, the model reduces output token usage by an average of 17% compared to its predecessor Gemini 3.5 Flash when completing the same tasks, with reductions reaching as high as 65% in software engineering benchmarks like DeepSWE. This means the model generates less redundant code and fewer loop operations when executing complex tasks, directly lowering enterprise API call costs.

On pricing, Gemini 3.6 Flash’s API is priced at $1.50 per million input tokens and $7.50 per million output tokens, down from the $9 output price of 3.5 Flash. Google emphasized that the simultaneous reduction in token consumption and unit pricing will significantly compress the total operational cost of individual AI agent tasks.

On performance, the model achieved quantifiable improvements across multiple benchmarks. In the DeepSWE test measuring software engineering capability, scores rose from 37% for 3.5 Flash to 49%; the machine learning engineering benchmark MLE-Bench jumped from 49.7% to 63.9%; the computer operation capability test OSWorld-Verified improved from 78.4% to 83%; and the knowledge work assessment GDPval-AA v2 rose from 1,349 points to 1,421 points.

Early customers including Figma, Harvey, Hebbia, and JetBrains gave positive feedback on Gemini 3.6 Flash’s balance of token efficiency, accuracy, and speed when handling complex workflows and knowledge-based tasks.

Lightweight Model Punches Above Its Weight, Speed and Cost-Effectiveness Stand Out

Gemini 3.5 Flash-Lite is positioned as the fastest and most cost-effective model in the 3.5 series. According to Artificial Analysis’s measured data, its generation speed reaches 350 output tokens per second, approximately double that of the previous-generation 3.1 Flash-Lite. Pricing is even more aggressive at just $0.30 per million input tokens and $2.50 per million output tokens, targeting high-throughput scenarios such as agent-based search and large-scale document processing.

Despite the “Lite” designation, the model outperformed the larger Gemini 3 Flash on multiple benchmarks. In the SWE-Bench Pro programming test, 3.5 Flash-Lite scored 54.2%, higher than 3 Flash’s 49.6%; in the OSWorld-Verified test, its 74.0% score also surpassed 3 Flash’s 65.1%. Developers can also dynamically adjust the model’s “thinking level” based on task complexity, flexibly balancing low cost and high quality.

Cybersecurity-Specific Model CodeMender: Autonomously Attacks to Verify Vulnerabilities, Automatically Generates Patches

The most eye-catching product of this release is undoubtedly Gemini 3.5 Flash Cyber and its deeply integrated code security agent, CodeMender. This model is purpose-built to discover, verify, and fix software vulnerabilities. Its defining feature: it not only statically scans code but also autonomously builds exploit verification code, simulating attacks in customer-managed isolated sandbox environments to confirm whether vulnerabilities genuinely exist and are exploitable.

Google noted that traditional static code pattern analysis tends to generate a high volume of false positives, leading to “alert fatigue” among security teams. Through dynamic attack verification, CodeMender effectively filters out false alarms, focusing efforts on genuinely dangerous vulnerabilities. Its detection scope covers complex vulnerabilities including memory corruption, injection attacks, cryptographic flaws, and insecure data handling, supporting mainstream programming languages such as C/C++, Go, Java, Python, Ruby, Rust, and TypeScript.

In the remediation phase, CodeMender automatically generates secure patches presented in diff format. After patch generation, the system’s built-in “LLM-as-a-judge” mechanism determines whether the fix breaks existing functionality. All patches must undergo manual review and approval by developers before being committed to code repositories, with administrative control always retained by the development team.

In the CyberGym security benchmark, CodeMender achieved results comparable to larger frontier models by calling up to five Gemini 3.5 Flash Cyber agents in parallel and integrating multiple analytical perspectives. In evaluations targeting the V8 JavaScript engine, the model discovered 55 unique and verified issues, 10 of which were missed by other models. Google has internally deployed CodeMender for vulnerability discovery and remediation in Chrome, Android, Google Cloud, Google Ads, and YouTube codebases. Its vulnerability research team once discovered a remote code execution vulnerability in a public API and a production memory corruption vulnerability within two hours.

Given the dual-use nature of cyber offense and defense technologies, Google has adopted an extremely cautious delivery strategy for Gemini 3.5 Flash Cyber. The model will not be open to all developers but is offered through CodeMender as a limited-access pilot exclusively to government agencies and trusted partners. Tulsi Doshi, Senior Director of Product Management for Google’s Gemini team, stated that the goal is to enable frontline defenders to patch critical vulnerabilities before they are widely exploited while reducing the risk of technology misuse.

Online brokerage Robinhood, an early CodeMender user, commented that the tool “accurately identified critical vulnerabilities that other AI tools completely missed.” Salesforce and Palo Alto Networks also gave positive evaluations.

Flagship Pro Absent, Gemini 4 Pre-Training Underway

In stark contrast to the three Flash models, the long-anticipated flagship model Gemini 3.5 Pro remains unreleased. Google’s last Pro series update was Gemini 3.1 Pro in February 2026. During this period, competitor OpenAI successively launched GPT-5.5 and began rolling out GPT-5.6, while Anthropic released Claude Opus 4.8 and Claude Sonnet 5.

Market sources indicate that Gemini 3.5 Pro’s code generation capability has consistently fallen short of internal performance targets. Google reportedly updated training data specifically to address shortcomings in late June, but with limited effect. There were even rumors of abandoning the original training plan and restarting from scratch. Logan Kilpatrick, technical lead at Google DeepMind, stated on social media only that 3.5 Pro is being tested with partners and will be widely available once ready, without providing a specific timeline.

Meanwhile, Google officially confirmed for the first time that it has begun pre-training work on “its most ambitious” Gemini 4, describing progress as exciting. This signals to some extent that Google’s model development pipeline continues to advance, but has also sparked speculation about whether the company is diverting attention from the flagship product’s delays.

Developer Community Feedback Divided, Cost-Effectiveness Questioned

After the new models became accessible, negative voices quickly emerged from the developer community. One user reported that when building a 3D scene, the model repeatedly ignored instructions after three prompts, ultimately producing output quality inferior to Gemini 3.1 Pro released five months earlier. Another developer pointed out that Gemini 3.6 Flash’s composite score on the Artificial Analysis platform is identical to 3.5 Flash, yet ranks behind multiple competitors including Meta Spark 1.1, GLM-5.2, GPT-5.6 Luna, and Claude Sonnet 5. Some users directly compared and noted that 3.6 Flash costs more to use than GPT-5.6 Sol medium while delivering lower intelligence, shaking the Flash series’ foundational value proposition of “high cost-effectiveness.”

However, some enterprise customers provided positive feedback. Robinhood, Salesforce, and Palo Alto Networks all expressed approval of CodeMender or Flash series models’ performance in specific scenarios, demonstrating that real-world application evaluations vary by use case.

All three new models became available on launch day via Google AI Studio, Android Studio, the Gemini app, and the Gemini Enterprise Agent Platform. Gemini 3.5 Flash-Lite will also be gradually integrated into Google Search. CodeMender is offered on an invitation basis through the Gemini Enterprise Agent Platform. Google plans to introduce third-party frontier models as options in the second half of 2026 and will soon enable integration with security firm Wiz, allowing direct invocation of CodeMender for code scanning from the Wiz platform.