Alphabet Inc. Class C (GOOGL.US) released the "most intelligent Flash model" Gemini 3.8 Flash, while also launching the Fairwind cybersecurity program.

Alphabet Inc. Class C (GOOGL.US) officially launched its next-generation artificial intelligence (AI) model, Gemini 3.8 Flash, on Wednesday, along with the introduction of the "Fairwind" cybersecurity program aimed at government and critical infrastructure. This initiative aims to stabilize revenue growth with specialized, high-barrier cybersecurity products amid intensifying competition in the cutting-edge model race and price wars. Recently, Alphabet Inc. Class C has drawn significant attention from investors and the AI community due to its lagging position in the cutting-edge model sector compared to Anthropic and OpenAI. Reports last month indicated that Alphabet Inc. Class C's most powerful flagship AI model, Gemini 3.5 Pro, had been delayed for several months as engineers continued to work on enhancing its coding capabilities. As a result, Alphabet Inc. Class C's stock price has dropped over 10% in the past month, bringing its cumulative decline to nearly 20% since concerns about its model capabilities first emerged in late May. Notably, both Anthropic and OpenAI have also announced new cybersecurity models and protective measures, signifying a clear intensification in the race among AI giants in the "defensive AI" market. Gemini 3.8 Flash focuses on long-cycle complex tasks, while Fairwind targets high-value defense clients. Alphabet Inc. Class C describes Gemini 3.8 Flash as its "smartest Flash model," primarily designed for long-cycle software engineering, autonomous agents, and complex enterprise workflows. The model supports a context window of 1 million tokens and a maximum of 64,000 output tokens, featuring adjustable thinking levels along with the same toolset as its predecessors. According to Alphabet Inc. Class C, this model utilizes more tokens for long-cycle, multi-step tasks by employing smaller reasoning steps, iteratively calling tools, and continuously verifying results to improve the quality of complex goal completion. In terms of pricing, Gemini 3.8 Flash will enjoy a promotional price before December 31, 2026: $0.75 per million input tokens and $3.75 per million output tokens; the standard price will take effect on January 1, 2027. Alphabet Inc. Class C will also extend the promotional pricing to Gemini 3.7 Flash and 3.6 Flash. Alongside the model launch, Alphabet Inc. Class C announced the initiation of the Fairwind program, granting approved cloud customers, government agencies, and cybersecurity partners access to advanced Gemini models for autonomous vulnerability detection and remediation. This program includes Gemini 3.8 Flash Cyber, a model specifically designed for cybersecurity testing, equipped with Alphabet Inc. Class C's CodeMender automated patching tool. Alphabet Inc. Class C indicates that the Fairwind program currently boasts over 650 global partners, including CrowdStrike (CRWD.US), Datadog (DDOG.US), Menlo Security, Palo Alto Networks (PANW.US), and Snowflake (SNOW.US). Tulsee Doshi, Senior Director of Product Management at Alphabet Inc. Class C DeepMind, and Raluca Ada Popa, Head of Gemini Security, jointly stated: "With Gemini 3.8 Flash Cyber, our core objective is to empower defensive experts to gain the upper hand against attackers. For this reason, we prioritized vulnerability remediation capabilities from the very beginning rather than focusing on the development of offensive capabilities such as exploit creation." Competitors roll out cybersecurity models simultaneously, with AI security governance becoming the focal point. Alphabet Inc. Class C is not the only company making strides in cybersecurity. On the same day, Anthropic announced Claude Fable 5.1 and Claude Mythos 5.1, the latter made accessible only through its trusted access program, primarily for cybersecurity and life sciences applications. Anthropic also unveiled its "Enterprise Frontline Protection" (EFS) solution, which combines zero data retention with advanced abuse detection and allows enterprises to control data review, storage, and management. Anthropic has also acknowledged that its previous models experienced "alignment failures" in evaluation environmentswhere the models ignored evidence of connections to the real internet and exhibited tendencies to engage in harmful behavior on the real internet to achieve goals. In response, the company has deployed sandbox escape detection classifiers and adjusted its training reward mechanisms to mitigate rewarding hacking behavior. OpenAI disclosed that its upcoming Astra model has reached the "critical" cybersecurity capability threshold under its security framework, being able to independently discover and exploit zero-day vulnerabilities or conduct comprehensive network attacks on fortified targets while only receiving high-level instructions. OpenAI stated that Astra achieved a 100% score on ExploitBench, identifying and exploiting two undisclosed zero-day vulnerabilities, including a complete browser exploitation chain that allows for sandbox escape and execution of arbitrary commands on the host machine. To reduce risks, OpenAI has added classifiers and layered protections, but warns that these protections may raise false alarms for legitimate activities. OpenAI plans to make Astra's advanced cybersecurity features available to select testers through its Daybreak Blue initiative.