The AI frontier just accelerated, faster, smarter, and more open than ever. MiniMax M2 redefines open-source intelligence with lightning-fast coding and agent performance at a fraction of the cost, Mistral AI Studio brings enterprise-grade observability and governance to production AI, and Anthropic supercharges its Claude models with a million Google TPUs, pushing the limits of large-scale training and deployment.

⚡ MiniMax M2: Efficient Open-Source AI for Coding & Agents – With a 230B MoE architecture and speeds up to 1500 tokens per second, M2 delivers frontier-level reasoning, coding, and autonomous agent performance at just 8% of Claude’s cost, free for a limited time.

🏢 Mistral AI Studio: Enterprise-Grade AI Platform – A full-stack production platform uniting observability, agent runtime, and governance so teams can trace, evaluate, and securely deploy AI systems with software-level rigor.

☁️ Anthropic Expands Google Cloud TPU Use – Anthropic taps up to one million TPUs worth tens of billions to train its next Claude models, scaling compute by over a gigawatt and solidifying its multi-platform AI infrastructure.

From open-source power and enterprise reliability to trillion-scale compute expansion, the next wave of AI is faster, more transparent, and built for everyone.

MiniMax M2: Efficient Open-Source AI for Coding & Agents

MiniMax M2 is an open-sourced AI model optimized for agents and code, offering top-tier coding and agentic performance at just 8% of Claude Sonnet's price and nearly twice its speed; it's available for free for a limited time, supports advanced tool use and deep search, powers a free general-purpose agent product with lightning-fast and pro modes, and can be self-hosted using released model weights with support for SGLang/vLLM, making it a cost-effective, high-performance solution for developers and businesses seeking affordable, rapid AI agent deployment. The model uses a sparse Mixture-of-Experts architecture with 230 billion total parameters (10 billion active per inference), achieves token speeds up to 1500/sec on modern GPUs, and scores in the global top 5 on comprehensive benchmarks for reasoning, coding, and agent orchestration. Its full-cycle code intelligence supports autonomous coding and debugging, and its robust automated agent workflows rival leading frontier models at a fraction of the cost.​

Mistral AI Studio: Enterprise-Grade AI Platform

Mistral AI Studio is a new production AI platform designed to help enterprise AI teams move beyond prototypes and build reliable, observable, and governed AI systems in production. It addresses common challenges such as tracking changes across model and prompt versions, reproducing results, collecting real usage feedback, running domain-specific evaluations, fine-tuning models privately, and deploying workflows with security and compliance. The platform is built around three core pillars: Observability, which offers tools to inspect traffic, build datasets, identify regressions, and trace outcomes back to prompts and usage; Agent Runtime, a durable and fault-tolerant execution environment for AI workflows that supports hybrid and self-hosted deployments; and AI Registry, a comprehensive asset management system tracking every AI lifecycle component with governance and versioning. Together, these enable enterprises to operate AI with software-level rigor, including continuous evaluation, traceability, governance, and flexible deployment. Mistral AI Studio aims to close the gap between experimentation and dependable AI operations, supporting secure, observable, and accountable AI deployments for large-scale business use.​

Anthropic Expands Google Cloud TPU Use

Anthropic is significantly expanding its use of Google Cloud’s TPU chips, gaining access to up to one million TPUs, valued at tens of billions of dollars, to dramatically boost its AI compute capacity. This expansion will bring over a gigawatt of capacity online by 2026, supporting the training and deployment of the next generation of Claude models. The decision reflects the strong price-performance and efficiency benefits seen with TPUs. Anthropic serves more than 300,000 business customers, and this additional capacity will help meet growing demand while enabling more thorough testing, alignment research, and responsible deployment. Anthropic's compute strategy employs a multi-platform approach using Google TPUs, Amazon Trainium, and NVIDIA GPUs, ensuring continuous advancement and strong industry partnerships.​

Hand Picked Video

In this video we’ll look at how to install and use Claude Code to build and ship a complete portfolio website from scratch, customize the design, and deploy it live with Vercel in minutes.

Top AI Products from this week

• Dynal.AI - Turn ideas, links, PDFs, and images into ready-to-post LinkedIn posts in minutes. Dynal learns your tone and uses proven patterns to draft editable copy and visuals, making it easier to post consistently and reach more people.

• Pokee AI - World's first vibe agentic workflow builder with an API that is as easy to use as ChatGPT. All Auth Handled. Agentic workflows now only require a single text prompt. No more manual node wiring. Reproducible. Reliable. Guaranteed.

• Grokipedia - Grokipedia is xAI's new AI-generated encyclopedia, powered by Grok. Pitched by Elon Musk as a more truthful alternative to Wikipedia, it launched with nearly 900k articles aiming to reshape online knowledge using AI.

• LLM Stats - LLM Stats is the go-to place to analyze and compare AI models across benchmarks, pricing and capabilities. Compare model performance easily through our playground and API that gives you access to hundreds of models at once.

• Namerly - Use AI to find the perfect, available brand name for your next project. Our AI brand consultant generates creative names and checks domain availability in real-time across multiple TLDs.

• Website Builder GPT - Generate fully functional, shareable websites directly inside ChatGPT — complete with instant live preview and one-click publishing. Perfect for rapid prototypes, landing pages, and personal projects.

This week in AI

• Reinforcement Learning Insights for Agentic Reasoning - This paper reveals key practices for enhancing agentic reasoning in LLMs via reinforcement learning, using real tool-use data, exploration-friendly techniques, and a deliberative strategy for better accuracy and efficiency.​

• Valthos Accelerates AI-Powered Biodefense - Valthos builds AI systems to identify and counter biological threats in real time, updating medical countermeasures instantly. Backed by $30M from OpenAI and top investors.

• Figma Buzz Updates - Figma Buzz now offers customizable design system-based templates, plugin integrations with tools like Brandfolder and Lokalise, and built-in video trimming, streamlining marketing and brand collaboration.

• AI Use Survey 2025 - 79.94% of people use AI personally or professionally. ChatGPT leads with 700M weekly users, followed by Meta AI (1B monthly) and Google Gemini (450M monthly).

• $100 ChatGPT Build - Andrej Karpathy shared a blueprint for building a ChatGPT-style AI for just $100, called NanoChat. It's a minimal, end-to-end, customizable pipeline for training and serving your own LLM.

Paper of The Day

This study explores how prompt politeness influences large language model (LLM) accuracy using 250 multiple-choice prompts with varying tones. Contrary to expectations, rude prompts consistently outperform polite ones, with accuracy ranging from 80.8% for very polite to 84.8% for very rude tones. This challenges earlier findings and suggests newer LLMs may prioritize core task understanding over politeness cues, raising important questions about prompting pragmatics and human-AI interaction.

To read the whole paper 👉️ here