In the week of October 10, China's "AI+" top-level design landed while frontier models raced globally. OpenAI released "dots", an all-weather agent built on GPT-6 Astra, equipped with an independent cloud computer and capable of maintaining task continuity across weeks. Google launched Gemini 4 Argon, first opened to cybersecurity defenders at $2 per million input tokens and $10 per million output tokens, alongside a general Gemini agent driven by goals rather than prompts that can run hours to days in the cloud across Gmail, Docs, Sheets, Drive, Chat, Calendar, Microsoft 365 and Slack.
Mistral published a preview of Mistral Large 4, a 1-trillion-parameter native multimodal MoE model activating only 49 billion parameters, trained on 3,800 NVIDIA Grace Blackwell GPUs in a European data center. Microsoft launched Microsoft-Decision-1, post-trained on the open-source Qwen3.5-9B for agent control, model routing and workflow decisions. Anthropic released Claude Haiku 5.5, its cheapest and fastest small model yet, with running costs about 75% lower than its predecessor and a 1M-token context window.
On the Chinese side, StepFun's Step 5 Preview topped OpenRouter trend charts a day after release, with open weights scheduled for October 15. Alibaba open-sourced Qwen-Image-2.1-Turbo, a 7B image generation and editing model that produces 2K images in just 8 denoising steps — five times fewer than the base model's 40 — and supports transparent-layer editing, making it one of the fastest models in its class. Chinese open-source models are accelerating their global expansion as inference costs fall.
Source: DAMO Developer Matrix / Official Announcements
