Google launches Gemini 3.8 Flash and specialized Cyber security model

Google claims its new Gemini 3.8 Flash model maintains the same introductory pricing as its predecessor, but the actual cost of using the tool has risen. While per-token rates remain at $0.75 per million input tokens and $3.75 per million output tokens, The Register reports that the model costs approximately 40% more per task than Gemini 3.7 Flash.

This price discrepancy stems from a design philosophy Google describes as "working harder." According to Google senior director of product management Tulsee Doshi and Gemini Security Lead Raluca Ada Popa, the model exhibits greater diligence on complex tasks by executing additional reasoning steps and calling tools iteratively. This allows the model to maximize performance, but it results in a 30% increase in output tokens per task and more turns during agentic evaluations, as noted by Artificial Analysis via The Verge.

Related YouTube video

The launch comes as Google attempts to stabilize its reputation following a period of leadership volatility and missed milestones. In August, Google DeepMind CEO Demis Hassabis moved to chairman, with Koray Kavukcuoglu taking over as SVP. These changes followed the company's failure to release Gemini 3.5 Pro in June as previously promised.

To manage these new costs, Google has provided an "escape hatch" for developers and enterprises prioritizing raw compute efficiency: they can either use lower effort settings or continue using Gemini 3.7 Flash, which remains fully supported.

Intelligence and Performance Benchmarks

Gemini 3.8 Flash is positioned as Google's most capable reasoning and coding model in the Flash line. On the Artificial Analysis Intelligence Index, the model scores 59 when set to high reasoning, a three-point increase over 3.7 Flash. This puts it on par with GPT-5.6 Sol (extra high) and Grok 4.6 (medium), though it still trails Claude Fable 5.1, which scores 66.

The model shows particular strength in specialized professional fields:

• Software Engineering: On the DeepSWE v1.1 benchmark, Google claims 3.8 Flash outperforms most larger frontier models in autonomously solving complex engineering problems end-to-end.
• Finance and Law: The model reported improvements on the Vals Finance Agent V2 and Harvey's Legal Agent benchmarks.
• Cost Efficiency: Artificial Analysis finds that at $0.58 per Intelligence Index task, Gemini 3.8 Flash is the cheapest model at its intelligence level, costing roughly six times less than Anthropic's Fable 5.1.

The Fairwind Program and Cyber-Specialization

Alongside the general model, Google has introduced Gemini 3.8 Flash Cyber. Unlike the standard version, which includes strict safeguards against misuse in cyber offense and Chemical, Biological, Radiological, and Nuclear (CBRN) domains, the Cyber variant features more permissive mitigations to assist defenders in finding and fixing vulnerabilities.

Because of these permissive settings, access is strictly gated through the new Fairwind Program. This program is limited to governments, software maintainers, and critical infrastructure operators. Its 650 members include the Center for Internet Security and CrowdStrike. Members gain access to the 3.8 Flash Cyber model and the CodeMender agent, designed to autonomously protect national security and public services.

On the CWE-Bench external benchmark, 3.8 Flash Cyber achieved a 47.2% pass@1 rate. This places it near the Pareto frontier, just behind a leading frontier model's 47.8%, while operating at a lower cost.

Competitive Landscape

The release coincides with a broader industry push toward specialized agents and cybersecurity capabilities. Rediff and CIO report that OpenAI's Astra model has reached a "Critical" cybersecurity capability tier, capable of discovering zero-day vulnerabilities. While Google's approach focuses on defensive patching through Fairwind, OpenAI's Astra has sparked debate over its "recurrent depth" architecture, which some researchers warn makes reasoning harder to monitor compared to traditional chain-of-thought processes.

Anthropic has also updated its offerings with Claude Fable 5.1 and Mythos 5.1, the latter of which is restricted to vetted users for biology and cybersecurity work. To compete on cost, Anthropic reduced prices for typical workloads by approximately 25%.

Google's rapid iteration—three Flash-tier releases in six weeks—suggests a pivot toward "agentic" workflows where the model operates more autonomously. To support this, Google also updated Gemini 3.7 Flash with agentic video understanding. This allows the model to scan video content dynamically rather than at a fixed frame rate, which Google claims reduces costs by up to 66% and token consumption by up to 88%.

The current financial incentive for developers to adopt these higher-reasoning workflows is temporary. The introductory pricing for Gemini 3.8 Flash is scheduled to double in the new year.

Deployment and Accessibility

Google has made Gemini 3.8 Flash available to a wide range of users across different platforms. Consumers with Google AI Pro or Ultra subscriptions can access the model through the Gemini app, AI Mode in Google Search, and Gemini in Google Sheets. For technical users, the model is accessible via the Gemini API in Google AI Studio and Android Studio, the interface design service Stitch, and Google Antigravity for developers, while enterprises can use Gemini Enterprise.

The deployment follows a period of accelerated iteration. According to The Register, Gemini 3.8 Flash is the fourth Flash model released by the company in four months. This aggressive cycle follows a June failure to release Gemini 3.5 Pro as promised.

Security Robustness and Utility

Beyond its reasoning capabilities, the 3.8 generation has seen specific technical refinements to its reliability. Google notes that the model has achieved measurable gains in robustness against prompt injection, a detail highlighted by Android Sage and verified by third-party evaluator Gray Swan. This is intended to protect agentic systems that must autonomously interact with untrusted tool calls and web content.

Google has also expanded its creative toolset with the launch of Google Pics, an image creation and editing tool. This tool is rolling out to Workspace business customers and Google AI Pro and Ultra subscribers, featuring integration into Google Slides and Google Docs.

The model's intelligence score of 59 places it in a competitive tier with Grok 4.6 (medium) and GPT-5.6 Sol (extra high). While it trails several flagship models in raw intelligence, Google's strategy relies on a cost-to-performance ratio that undercuts competitors, provided users manage the increased token consumption resulting from the model's iterative reasoning process.

For developers, the next step involves preparing for a significant cost shift, as the current introductory pricing is scheduled to double in the new year.