This week's technical AI news: OpenAI cuts prices across its GPT-5.6 lineup, Microsoft's narrow security specialist model outperforms bigger frontier competitors, and Google's open-weight Gemma family crosses a major adoption milestone.

1. OpenAI cut prices on its GPT-5.6 model family on July 30, with the budget-tier Luna model dropping 80 percent to $0.20 per million input tokens and $1.20 per million output tokens, while the mid-tier Terra fell 20 percent to $2 per million input tokens and $12 per million output tokens. The flagship Sol model kept its $5/$30 pricing but gained a new "Fast" mode running up to 2.5x faster at double the cost. OpenAI says the cuts stem from efficiency work in which Sol itself helped optimize the infrastructure serving it -- a self-improving efficiency loop that let the company pass savings straight through to API, ChatGPT Work, and Codex customers. The move intensifies an industry-wide price war as rivals increasingly compete on cost rather than raw capability. https://openai.com/index/advancing-the-price-performance-frontier-with-gpt-5-6/

2. Microsoft unveiled MAI-Cyber-1-Flash, its first in-house cybersecurity AI model, built specifically for vulnerability discovery rather than general-purpose reasoning. Paired with GPT-5.4 inside Microsoft's new MDASH multi-agent system, the combination scored 95.95 percent on the CyberGym benchmark -- more than 10 percentage points ahead of comparable setups from Google, OpenAI, and Anthropic. The specialist model handles roughly 90 percent of routine security tasks on its own, escalating only the hardest cases to a larger model, which Microsoft says cuts operating costs roughly in half versus prior configurations. The broader platform, Project Perception, entered public preview on August 3 for approved customers on Azure AI Foundry, signaling a bet that narrow, cheap specialist models can beat frontier generalists at well-defined tasks. https://www.securityweek.com/microsoft-unveils-mai-cyber-1-flash-its-first-cybersecurity-ai-model/

3. Google's open-weight Gemma model family crossed 900 million cumulative downloads, with Gemma 4 -- released in April -- alone accounting for more than 300 million of them. The milestone underscores how quickly open models have moved from research curiosities to production infrastructure: developers have published more than 70,000 fine-tuned Gemma variants on Hugging Face, spanning lightweight edge deployments to larger multimodal reasoning models. The numbers arrive as open-weight competition intensifies from Chinese labs like Moonshot AI and Z.AI, making Gemma's adoption curve a proxy for how much of the open-model market Google is capturing. https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/