TLDR: This week brought faster and cheaper coding models, renewed momentum for open weights, frontier cyber capabilities moving into enterprise clouds, and another potential multibillion-dollar AI acquisition.
Google Launches Gemini 3.7 Flash for Coding and Agent Workflows
On August 13, Google launched Gemini 3.7 Flash, a lower-cost model aimed at software coding and autonomous business workflows. Google says the model improves debugging, issue resolution, and production-ready code generation while giving agents stronger ability to plan tasks, use tools, and complete multi-step work with less human intervention. To drive adoption, Google priced it at an introductory $0.75 per million input tokens and $3.75 per million output tokens through the end of 2026, half the original price of Gemini 3.6 Flash.
OpenAI Puts GPT-5.6 Sol Into Ultrafast Mode at Up to 14X the Speed
On August 13, OpenAI previewed Ultrafast, a new API service tier that runs GPT-5.6 Sol up to 14 times faster than standard processing. Powered by Cerebras, the system can generate up to 750 output tokens per second while keeping the intelligence of OpenAI's flagship model, making it practical for time-sensitive work such as incident response, live financial research, customer support, voice systems, and interactive experimentation. Ultrafast is starting as a limited preview for selected customers, with wider access planned as capacity grows.
Meta Returns to Open-Weight AI With Muse Glimmer and Promises Bigger Models
On August 10, Meta released Muse Glimmer, a compact open-weight model designed to run agentic tasks on a Mac or PC using a single graphics card. Mark Zuckerberg paired the launch with a broad argument for lowering U.S. barriers around open AI and said Meta plans to release the weights of the more powerful Muse Spark 1.2 next. The company also announced a $1 billion fund for communities affected by data-center construction, tying its return to open models to a wider push for cheaper AI, local deployment, and faster U.S. infrastructure growth.
China's GLM-5.3 Comes Within Reach of Mythos 5 on Cyber Defense
On August 14, Chinese AI startup Z.ai said its upcoming GLM-5.3 model scored 84.5% on CyberGym for finding and confirming software vulnerabilities, slightly above the 83.8% result it reported for Anthropic's restricted Mythos 5. The gap widened on exploit development, where GLM-5.3 scored 54.4% against Mythos 5's 78.0%. Z.ai says it will hold the model for roughly two weeks of additional safety testing before public release, while keeping its most sensitive cybersecurity capabilities behind a verified-user trusted-access program.
Nvidia Builds a 1-Trillion-Parameter Nemotron 4 as Open Models Heat Up
On August 11, Reuters reported that Nvidia is developing Nemotron 4, a new open-model family whose largest version is expected to contain at least 1 trillion parameters. Final training is not yet complete, but employees cited by The Information said the system could be ready as early as late fall. Nvidia also released Nemotron 3.5 Lightning for tasks such as code review, tool use, and security monitoring, alongside NeMo Switchyard, an open-source router that automatically sends each AI task to the model best suited to handle it.
Anthropic Explores a $6 Billion Decart AI Deal Ahead of Its IPO
On August 13, Reuters reported that Anthropic is in talks to acquire Nvidia-backed startup Decart AI as the Claude maker looks for more infrastructure and optimization capacity ahead of a potential public listing. Bloomberg reported that a deal could be worth around $6 billion. Decart develops AI infrastructure alongside models such as Lucy, which edits live video in real time, and Oasis, which generates simulated environments for robotics and autonomous-driving research. If completed, Decart's team would reportedly join Anthropic's inference and performance organization.
OpenAI Brings Daybreak Red and Blue Cyber Models Into Amazon Bedrock
On August 11, OpenAI made its Daybreak cybersecurity capabilities available through Amazon Bedrock, giving approved AWS customers a direct path to use frontier cyber models inside existing enterprise environments. Daybreak Blue provides GPT-5.6 Sol and other general-purpose frontier models with safeguards for authorized defensive work, while Daybreak Red provides purpose-trained models for advanced vulnerability research, exploit validation, and security testing. The rollout pushes OpenAI's most sensitive cyber capabilities deeper into mainstream cloud security workflows without requiring teams to build a separate AI infrastructure stack.
