THE BIG ONE
Deepseek makes a bold move
Deepseek has made its 75% discount on the V4-Pro model permanent, pricing output tokens at $0.435 per million input tokens. This positions it at 11.5 times cheaper than GPT-5.5 and over 34 times cheaper on output tokens. This pricing strategy is likely to disrupt the competitive landscape, making advanced AI models accessible for smaller companies and startups. With operating costs significantly reduced, developers can experiment more freely and innovate without the hefty price tag. However, it raises questions about the sustainability of such pricing in the long run. As builders, keeping an eye on how this affects the market dynamics will be crucial. Read more here.
QUICK HITS
Alibaba's Qwen3.7-Max runs autonomously: Alibaba's new model, Qwen3.7-Max, has demonstrated impressive capabilities by autonomously optimizing code for its custom chip over 35 hours. This showcases the potential for AI agents to handle complex tasks without constant oversight. Why it matters: It could redefine how we approach long-running tasks in production. Read more.
UC Berkeley Law bans AI in education: Starting summer 2026, UC Berkeley Law will restrict AI usage in nearly all graded work, allowing only research use. This decision emphasizes the need for human oversight in fields that demand critical thinking and ethical considerations. Why it matters: It could set a precedent for other institutions, influencing how AI is integrated into higher education. Learn more here.
Microsoft's Fara1.5 outperforms competitors: Microsoft Research has released Fara1.5, a family of browser computer-use agents that outperformed OpenAI's Operator and Gemini 2.5 in benchmarks. This release indicates how competition is heating up among AI models, pushing the boundaries of performance and efficiency. Why it matters: It highlights the rapid advancement in agent architectures. Read more here.
OpenAI's new PowerPoint plugin: OpenAI has launched a ChatGPT PowerPoint plugin that creates presentations from various inputs. However, users are cautioned that the plugin could accidentally delete content. Why it matters: While this feature enhances productivity, it also underscores the need for careful integration of AI tools in creative workflows. Get the details here.
Cohere's Command A+ model: Cohere's new Command A+ model is a 218B Sparse Mixture-of-Experts model that can run efficiently on minimal hardware. This is a significant step for democratizing access to powerful AI tools. Why it matters: It allows more developers to experiment with advanced AI without needing extensive resources. Learn more about it here.
ONE THING TO TRY
If you’re looking to experiment with AI agents, try implementing GBrain, the self-wiring memory layer developed by Y Combinator’s Garry Tan. It allows agents to retain context across sessions, enhancing their utility in long-term projects. It’s open-source and can provide a robust solution for managing memory in your AI deployments.
SIGN-OFF
That’s it for this week’s insights! I hope you found something useful to take away and apply in your own projects. If you have any thoughts or experiences with these tools, I’d love to hear from you!