Google has once again made a move in the competitive field of artificial intelligence with the launch of Gemini 3.8 Flash and Gemini 3.8 Flash Cyber, two models that arrive just weeks after the unveiling of Gemini 3.7 Flash. The American company seeks to solidify its Flash family as a balanced option in terms of speed, cost, and advanced capabilities, especially in the areas of software development and the automation of complex tasks.
The new general-purpose version promises significant advancements in code engineering and agent-based workflow management, while the Cyber variant is geared towards cyber defense, with the ability to locate security vulnerabilities and automatically generate patches. Both models maintain the low-cost philosophy that has characterized the Flash series, although access to the security version will be restricted through a specific program.
Gemini 3.8 Flash: more intelligence without raising the price
The standard model for this new batch is presented as the smartest Flash Google has created to date. The main improvement focuses on code generation and understanding, as well as agentic workflows—processes in which a system breaks down a complex task into smaller steps, consults tools or databases, and completes the operation with minimal human intervention. This is especially useful for software development, business automation, and document analysis.
In internal testing conducted by the company, Gemini 3.8 Flash outperformed its predecessor in specialized reasoning and programming tasks. For example, in the DeepSWE v1.1 benchmark, which assesses the ability to solve complex and long-range tasks, the new model outperformed larger and more expensive competitors, including some of the most advanced systems from OpenAI and Anthropic. It also performed well in domains such as finance and law, where it surpassed the previous version in specific analysis and reporting tests.
One of the keys to this improvement lies in the model's ability to adjust its effort according to the complexity of the problem. When the task requires it, Gemini 3.8 Flash executes additional reasoning steps and uses tools iteratively, allowing it to tackle more difficult tasks without sacrificing speed on simpler queries. This approach, which Google calls "adaptive effort," represents a shift from the traditional dynamics of AI models, which typically apply the same level of computation to all requests.
Regarding pricing, the company has decided to maintain the launch fee for Gemini 3.7 Flash: $0,75 per million entry tokens and $3,75 per million exit tokens. This decision aims to ensure that the increased capabilities do not result in higher costs for developers and companies already using these models in their applications and services.
Availability of tools for developers and users
The Gemini 3.8 Flash rollout covers a wide range of products within the Google ecosystem. Developers will be able to access the model through Google AI Studio and Android Studio, as well as through Google Antigravity, the company's environment for agent flows. Organizations will also find it in Gemini Enterprise, while Google AI Pro and Ultra subscribers will have access from the Gemini app on Android.
Furthermore, the model is now available in Google Search's AI Mode and Gemini in Google Sheets. This means users can use it to write or review code, research queries, work with spreadsheets, or automate internal processes directly from these tools. The integration with the search engine is particularly relevant, as AI Mode increasingly relies on Gemini to answer search queries.
Google hasn't specified whether access will be enabled simultaneously in all countries, languages, and supported accounts, but the company has confirmed that the feature has been available on major platforms since the announcement. Users with free accounts will have to wait, as access through the Gemini app and AI Mode is reserved for paid subscribers.
Gemini 3.8 Flash Cyber: the bet on automated cyber defense
The biggest news in this announcement is undoubtedly Gemini 3.8 Flash Cyber. Google presents it as its most advanced cyber defense model to date, designed specifically to address issues related to computer security. Its main objective is to detect software vulnerabilities and help develop patches automatically, a task that traditionally requires weeks or even months of manual work.
In tests conducted by the company, Flash Cyber has demonstrated remarkable performance. In the CWE-Bench benchmark, administered by Collinear, the model achieved a first-attempt success rate of 47,2%, very close to the 47,8% achieved by the leading model in the field, but at a significantly lower cost. It also exceeded 70% success in detecting vulnerabilities across a catalog of 20 different programming languages, according to an internal assessment.
The results in real-world environments are equally striking. Chrome's security team has found that Gemini 3.8 Flash Cyber generates 2,6 times more correct patches than other, larger commercial models. Meanwhile, Google's cloud vulnerability research team used this AI to find a critical vulnerability in under two hours, when such research typically takes months. The security firm Wiz has also reported a detection rate 7,5% to 9,7% higher in its internal penetration tests, at a cost 2,3 to 5,2 times lower than other advanced models.
However, the offensive potential of this tool is clear. A system trained to understand and repair a security flaw also knows how to exploit it. For this reason, Google has decided to restrict access to Flash Cyber through the new Fairwind program, which limits its use to trusted stakeholders, such as government agencies, critical infrastructure operators, and software maintenance providers. This measure follows similar initiatives by OpenAI, Anthropic, and Microsoft, which have also launched restricted access programs for their cybersecurity models.
A launch that sets the pace for the Flash family
The arrival of Gemini 3.8 Flash comes at a time of intense activity for Google in the field of artificial intelligence. The company recently reorganized the leadership of Google DeepMind, with Demis Hassabis assuming the role of president and chief scientist, while Koray Kavukcuoglu has taken over as head of the unit. It has also brought on Barret Zoph, formerly head of post-training at OpenAI, as vice president of research, tasked with driving reinforcement learning and post-training of the models.
These moves reflect the pressure Google feels to regain ground against rivals like OpenAI and Anthropic, which have gained traction with tools capable of writing, debugging, and running code more autonomously. The Flash family has become Google's answer to this competition, offering faster and more affordable models than the flagship versions, but with sufficient performance for the typical tasks of development teams.
The launch of Gemini 3.8 Flash and its Cyber variant demonstrates that the company is committed to specialization and efficiency as a strategy to differentiate itself in an increasingly saturated market. While other manufacturers compete to offer the largest and most powerful model, Google seems to have found a niche in combining speed, cost, and specific capabilities for particular sectors.
With this third Flash release in just six weeks, Google is sending a clear signal to the market: artificial intelligence is no longer measured solely by raw power, but also by its ability to adapt to users' real needs. The technology is beginning to take control of its own review and correction processes, becoming autonomous agents capable of executing and auditing complex tasks within the same digital ecosystem in which they operate.
The coming months will be crucial in determining whether Gemini 3.8 Flash lives up to expectations and whether the Cyber variant establishes itself as a leading tool in the cybersecurity field. For now, Google has successfully presented a robust offering that combines innovation, efficiency, and responsible use of a technology with a clear dual-purpose nature.
