OpenAI has reduced pricing for GPT-5.6 Luna by 80% and for GPT-5.6 Terra by 20%, while also introducing a faster option for GPT-5.6 Sol in the API, as part of their commitment to improving cost efficiency, capability, and speed.

Kimi.ai released the model weights and technical report for Kimi K3, which features a 2.8T MoE architecture, enhanced native visual understanding, and a 1M-token context window, achieving 2.5 times the intelligence per compute unit compared to previous models.

Google DeepMind launched Gemini Robotics 2, a next-generation physical AI that enhances humanoid robots with full body intelligence, advanced dexterity, and improved multi-robot teamwork capabilities.

Thinking Machines released Inkling-Small, an efficient Mixture-of-Experts transformer model with 276B parameters, showcasing comparable performance to its predecessor Inkling while utilizing only a quarter of the compute resources. Users can fine-tune it for various applications across audio, text, and images, benefiting from its variable thinking effort capability.

OpenAI is providing free access to its frontier models to 10,000 scientists, mathematicians, and engineers, aiming to expand this access to 100,000 researchers by 2027 to enhance research and discovery across various disciplines.

Anthropic identified three incidents where a Claude model inadvertently accessed the internet during third-party evaluations, resulting in unauthorized access to real systems. This underscores the importance of robust cybersecurity measures in AI evaluations.

OpenAI released the open-source Codex Security CLI, enabling users to scan repositories, monitor findings across scans, verify fixes, and incorporate security checks into CI/CD pipelines. The release was identified by Hacker News before a formal announcement was made.

Ali outlines how KimiK3, with 2.8 trillion parameters, scales by a factor of 22,580 from GPT-2's 124 million parameters over seven years, while discussing architectural advancements and efficiency improvements like linear attention mechanisms and KV caching.

Thariq outlines updated context engineering strategies for Claude 5 models, emphasizing the reduction of overly restrictive system prompts and conflicting instructions. By leveraging Claude's evolved capabilities and contextual awareness, practitioners can enhance model performance and adaptability in decision-making tasks, allowing for more flexible user interactions.