The briefing · 23 Sep 2026 · 72 tracked changes
OpenAI Advances Agent Infrastructure While Global AI Tools Expand Integrations
Today’s developments highlight significant updates to OpenAI’s agent capabilities, price cuts for developers, and key hardware optimizations from major chipmakers.
OpenAI Powers Up Agent with GPT-6 and MCP Server
OpenAI has upgraded its ChatGPT Agent to run on its new GPT-6 model stack, which includes the GPT-6 Astra version. Additionally, the agent now features a Model Context Protocol (MCP) server, enhancing its ability to connect with external data and tools. These updates point toward more autonomous and capable software agents that can perform tasks directly on computers.
Why it matters Advanced AI agents will soon handle complex, multi-step digital chores on your behalf by integrating directly with your existing software.
Built from 3 tracked changes
Developer and Design Tools Expand Access
Anthropic has updated Claude Code, adding a command-line interface, a JetBrains plugin, and a web interface, alongside reducing its entry price to $17 per month. Meanwhile, Cognition has equipped its software engineering agent, Devin, with a new SWE-2 model and shell access. In the creative design space, Figma dramatically lowered its AI entry price to $3 per month, and Canva integrated the Model Context Protocol.
Why it matters Lower pricing and wider software integrations make highly specialized coding and design AI tools accessible to a broader range of professionals.
Built from 5 tracked changes
Asian Language Models Level Up on Reasoning and Context
Baidu has introduced ERNIE 5.0 (Thinking Preview) with a expanded 128k context window, designed for more deliberate reasoning. Tencent updated its Hunyuan model to the 'Hunyuan-role-latest' version, introducing speech-to-text, text-to-speech, and an OpenAI-compatible API. MiniMax has also expanded its chat platform's context window capacity to 1 million tokens, allowing it to process massive volumes of text in a single request.
Why it matters International AI models are becoming highly competitive, offering larger memory spaces and deeper thinking capabilities for global enterprises.
Built from 5 tracked changes
Next-Gen Chips Optimize for Efficiency and Speed
NVIDIA Blackwell added hardware support for FP4 AI Inference and FP8 Precision Training, boosting calculation efficiency. Tenstorrent transitioned its Blackhole chip to its production version, introducing liquid-cooled support, a low-level C++ SDK, and an open-source compiler. Google also moved its Ironwood TPU to its 7th generation architecture, maintaining the intense pace of the hardware infrastructure race.
Why it matters More efficient, liquid-cooled, and precision-optimized hardware is crucial to lowering the high energy costs and physical footprints of running advanced AI.
Built from 5 tracked changes