Google Releases Three Models Including Gemini 3.6 Flash — Pre-training for Gemini 4 Also Underway

On July 21, Google rolled out three Gemini Flash series models simultaneously. The flagship **Gemini 3.6 Flash** improves coding, knowledge work, and multimodal performance while cutting output tokens by up to 17% compared with 3.5 Flash (up to 65% on the DeepSWE benchmark). It aims to be the "workhorse" that supports developer productivity on the cost-performance front.

**Gemini 3.5 Flash-Lite** is the fastest and cheapest model in the 3.5 class, generating 350 tokens per second according to the Artificial Analysis Index. It also delivers improved performance in agentic workflows compared with previous Flash-Lite generations.

Meanwhile, **Gemini 3.5 Flash Cyber** is a lightweight model specialized in code security, offered exclusively to governments and trusted partners in combination with the CodeMender agent. The rationale: when scanning large codebases for vulnerabilities, repeatedly running a lightweight specialized model is more efficient than repeatedly invoking expensive large models.

Notably, the frontier model **Gemini 3.5 Pro remains in partner testing** and has not been made broadly available. The Decoder observes that "while OpenAI, Anthropic, and Chinese labs are already competing at the frontier level, Google's flagship model remains adrift." At the same time, Google has already kicked off **its most ambitious pre-training run yet for Gemini 4**, laying the groundwork for the next generation.

Microsoft and Mistral Expand Multi-Billion-Dollar Partnership for European AI Infrastructure

Microsoft and France's Mistral have significantly expanded their strategic partnership, signing a multi-billion-dollar agreement to build AI infrastructure across Europe. In the context of securing European AI sovereignty, the push to reduce dependence on US cloud infrastructure is accelerating. Detailed investment amounts and data center scale have yet to be announced, but the deal is expected to form the foundation for deploying Mistral's models to European businesses and public institutions.

OpenAI's Agent Hits 10 Million Users After Its ChatGPT Work Debut

According to Bloomberg, OpenAI's agent feature **rapidly reached 10 million users following the launch of ChatGPT Work**. The figure signals that agentic capabilities are entering the practical stage for everyday users as well. OpenAI is accelerating its strategy of building agent features into business workflows, and the trajectory of its user base will be watched as a growth indicator for the emerging agent economy.

US Treasury Secretary Hints at Sanctions Over "IP Theft" by Chinese AI Models

On July 21, Treasury Secretary Scott Bessent raised the possibility of **imposing sanctions**, citing intellectual property infringement by Chinese open AI models. The move is part of the Trump administration's campaign to curb Chinese AI, and it appears to open a new phase targeting the outflow channels of open-source models. According to CNBC and TechCrunch, the models in scope and the specific sanctions measures remain unclear, but they could affect the use and distribution of open-weight models.

Altman to Brief Senior US Government Officials on Next-Generation AI Models Next Week

Also per Bloomberg, OpenAI CEO Sam Altman is scheduled to **brief officials from the Trump administration and Congress on the next-generation AI model family** next week. The move is tied to the AI safety review framework the US government has in development (expected to be finalized within weeks), with the models' capabilities and their impact on employment among the topics to be covered. If the safety review becomes institutionalized, government involvement would be built into the release process for frontier models themselves.

New Malware Takes Aim at AI Infrastructure — Wipes Data with a "Kill Switch"

Worth noting is a new type of malware targeting AI coding systems, reported by WIRED on July 21. The tool reportedly burrows deep into AI development environments, and beyond stealing data and login credentials, it can **destroy files by flipping a "kill switch"** and lock out legitimate users. The arrival of this new breed of cyberattack targeting AI infrastructure once again underscores the importance of security measures.

---