爱窝啦 AI 日报

Today’s Summary

Xiaomi's MiMo-V2.6 Pro tops the open-weight leaderboard and open-sources 7,000 training environments, while Anthropic and OpenAI both drop prices and launch new models on the same day.
From open-source frameworks to office suites, Agents are evolving from research tools into products that can be directly installed into systems and workflows.
Start with the pricing and open-source stories in the spotlight section, then try Bob Translator's custom Prompt feature.

🔥 Today’s Top 10 Highlights

1. Xiaomi Open-Sources MiMo-V2.6 and Tops the Open-Weight Leaderboard

Pro version claims the top spot. According to 36Kr, Xiaomi open-sources MiMo-V2.6 and tops the open-weight leaderboard . Xiaomi simultaneously released around 7,000 training task environments, RL version weights, and supporting frameworks. Reinforcement learning researcher Nathan Lambert praised it as “impressive.” A desktop client for everyday users launched at the same time.

2. Agent Substrate System Hits GitHub Daily Trending

Multi-agent coordination framework gains traction. GitHub Trending Daily shows Agent Substrate system hits GitHub daily trending , reaching 2,976 total stars. The project implements core systems in Go, designed specifically for multi-agent parallel tasks. Good fit for teams building Agent infrastructure.

3. Codex Now Supports GPT-6 Sol and Luna

Two fast models available. A V2EX user post shows GPT-6 Sol is now live on Codex , alongside GPT-6 Luna. The screenshot confirms both new models appear in the selector. Perfect for Codex Plus subscribers who need faster inference speeds.

4. Alibaba Cloud Sponsors Omarchy Foundation and Launches Localization

Easier installation in China. Guizang points out in a tweet that Alibaba Cloud sponsors Omarchy Foundation and launches localization , and will launch Omarchy China. Omarchy will also be integrated into the newly released Qwen Book. DHH emphasized in the announcement that “agentic computers need native agentic operating systems.” Local developers can expect a smoother installation experience.

5. Today AI China Version Fully Launches September 24

Personal assistant that proactively helps you. After beta testing, Guizang says Today AI officially launches in China on September 24 , calling it the first domestic product he’s seen that truly embodies the personal agent concept. Today gradually learns user-related information and proactively assists when needed. Supports iOS, Android, Windows, Mac, Linux across all platforms.

6. Bob Translation Tool Supports Custom Prompts

One-time purchase of 50 yuan, customize prompts freely. Xiangyangqiaomu tested and reported that Bob translation tool has built-in free models from Zhipu and Silicon Flow , and supports custom prompt mode. Users can copy configurations and modify instructions, like “explain in simple and concise terms: $query.text.” Select any text and press option + d to trigger custom interpretation.

7. Upscayl Open-Source Image Super-Resolution Tool

Runs locally without uploading to servers. Telegram channel aigc1024 recommends Upscayl specifically for rescuing blurry images and old photos . Uses AI super-resolution models to upscale images and fill in details, supports old photos, anime images, wallpapers. Completely runs locally, photos never pass through servers. Supports Windows, macOS, Linux.

8. Anthropic Releases Opus 5.5 with 40% Price Cut

API costs down by 40%. Anthropic’s official tweet shows Claude Opus 5.5 costs 40% less than Opus 5 , with cache reads down 60%. The new model performs close to Claude Fable 5.1 on most tasks. Now available in the API. Teams running large-scale inference tasks can switch models today and recalculate costs.

9. OpenAI Launches GPT-6 Sol and GPT-6 Luna

Two faster models go live. Sam Altman retweeted the official announcement that OpenAI launches GPT-6 Sol and GPT-6 Luna , with speed and pricing better suited for scaled deployment. OpenAI also optimized caching and inference efficiency. Developers can reduce bulk API costs while maintaining capabilities.

10. Yandex Open-Sources 80B MoE Large Model

Russia enters the large-parameter race. Gorden Sun introduces in a tweet that Yandex open-sources 80B MoE large model , Yandex’s first open-sourced large-parameter model. Total parameters 80B, activated parameters 3B, MoE architecture trained from scratch. Capabilities currently fall short of DeepSeek Flash 0731.

⚡ Product and Feature Updates

Qianwen Launches Independent Office Suite Product

Not Qianwen Office. Telegram channel aigc1024 mentions Qianwen launches an independent office product called “Qianwen” , different from the previous Qianwen Office. Specific features and availability haven’t been fully disclosed yet. Users interested in AI tools for office scenarios can watch for updates.

⌘ Top Open-Source Projects

dream-num/univer: Office Runtime for AI Agents

Integrates five document types. GitHub Trending Daily shows univer gained 255 stars that day , reaching 15,400 total stars. The project implements spreadsheets, documents, slides, canvases, relational tables, and PDFs in TypeScript, all running in a unified AI Agent environment.

davila7/claude-code-templates: Claude Code Configuration and Monitoring CLI

Command-line management for dev environments. GitHub Trending Daily shows claude-code-templates gained 64 stars that day , reaching 31,122 total stars. The tool implements a CLI in Python, specifically for configuring and monitoring Claude Code development workflows.

◉ Social Media Highlights

Baoyu Retweets Astra and Tripo Recreating Titanic Interactive Movie

Complete movie generated with five minutes of code. Baoyu retweeted Jared Zongyuan’s test showing Astra plus Tripo recreated the Titanic interactive movie . The original author believes Opus 5.2 has reached AGI level, capable of generating a five-minute complete film with code and showing the entire sequence of events.

Xiangyangqiaomu Tests Claude Opus 5.5 Across Benchmarks

Improvements across Terminal, Knowledge, and Operation leaderboards. Xiangyangqiaomu tested and found Claude Opus 5.5 shows major improvements on Terminal-Bench 4.0, GDPval-AA v2.1, and OSWorld 2.0 benchmarks . Speed up 30%, cost down 40%. They also tested pelican-on-bicycle generation, calling it the best result they’ve seen yet.

😄 AI Fun

New US Account Owner Worries About Limits Right After Activation

A user posted on V2EX asking what’s the normal 5-hour limit for a newly opened US Codex Plus account . The post included a screenshot of the limits, with the account just activated and already researching whether the quota is enough. GPT-6 Sol just launched, and users are already worried about being flagged as abnormal accounts by the system. Haven’t even started using it, but already figured out all the support issues.

❓ Related Questions

How to use GPT-6 Sol and Luna in China?

GPT-6 Sol and Luna are now live on OpenAI Codex Plus. According to V2EX user posts and Sam Altman’s retweet of the official announcement, these two models inherit GPT-6 Astra’s core capabilities, with speed and pricing better suited for scaled deployment.

Currently requires an active Codex Plus subscription to use. Regional support, payment methods, and quota limits are subject to OpenAI’s official real-time pages.

For further comparison of currently available AI accounts or subscription services, check Aivora AI Account Store ; products, prices, and availability are subject to the official website’s real-time pages.