According to reports, Google acquired Spirit Airlines' emails, Teams chat, and other business data for $10 million in a bankruptcy auction, and plans to use it for product development and AI model training. The transaction still requires court approval.
Both OpenRouter and Vercel AI Gateway are currently showing a 50% discount on GPT-5.6 Sol; Vercel explicitly states that the promotion lasts until September 18, 2026.
SpaceX has officially acquired Grok for $60 billion, and the team will join SpaceXAI to jointly develop the Grok series and Cursor products. This is the largest startup acquisition in history. Cursor will gain stronger computing power by leveraging SpaceX's world's largest GPU cluster. At the same time, xAI released a new AI companion product, Grok Bot, which can autonomously run multi-step real-world tasks 24/7.
The Dots API Open Platform provides the dots3-note-prev calling interface, which is compatible with the OpenAI Chat Completions format and supports 512K context, multimodal input, tool calls, and streaming responses.
Xiaohongshu's Dots Model Lab has open-sourced the Dots3-note preview, the first open-source version of the Dots3 series. The model has a total of 280 bytes of parameters and 16 bytes of activation parameters, supports up to 512K of context, and possesses multimodal understanding capabilities including text, vision, and speech. It is optimized for complex reasoning, agent-based processing, and multimodal perception. The model performs comparably to larger models with several times the number of parameters on multiple benchmarks.
MiniMax-Music3, an open-source music model, supports lyrics generation for songs up to 5 minutes long. The model employs a layered architecture: an 8-byte global model handles long-range semantics and structure, while a 0.6-byte local model recovers fine-grained acoustic information. MiniMax-Music3 outputs 32kHz stereo WAV, preserving theme, rhythm, vocal identity, and arrangement progression, covering intros, verses, choruses, and other structural elements.
Zhipu launched GLM-5.3, maintaining the same foundation but significantly improving the upper bound of intelligence through training scaling, making it the most programmable open-source model currently available. It achieved first place in multiple benchmark tests, including Terminal Bench 3.0. The model also demonstrates strong cybersecurity capabilities, excelling in tasks such as code review and vulnerability discovery. Zhipu, in collaboration with multiple security teams, has discovered a total of 2436 vulnerabilities and launched the "Open Source Shield" project.
Google has launched the Gemini 3.7 Flash model, designed specifically for programming and agent scenarios. The model offers significant improvements in software engineering, knowledge work, and web development. From now until the end of 2026, the initial launch price is half the original price of the 3.6 Flash model: $0.75 per million tokens input, $3.75 output. Gemini Spark has integrated this model, optimizing its Workspace tool capabilities.
DeepSeek Harness Developer Preview (v0.1) is open for testing by Harness developers worldwide and is also open-source under the MIT license. The product adopts an "everything is a plugin" architecture, built on the Cordis plugin system, allowing for free replacement and combination of Agent capabilities such as models, tools, and skills. It provides four operating modes: Standard, PTC, Minimalist, and Creative, supporting full tool calls and custom presets. Harness Team
Meoo has launched a memory feature that automatically organizes and saves long-term, valid information such as a user's personal background, work habits, response style, and project rules, which can then be intelligently recalled in subsequent conversations. Memory is divided into "personalized memories" visible only to the user, including user profiles, work manuals, response styles, memory fragments, and application memories shared by the team.
The Xiaohongshu Dots team, in collaboration with the X-LANCE Lab at Shanghai Jiao Tong University, open-sourced dots.tts, a fully continuous latent space autoregressive TTS foundation model with 2 billion parameters. The model does not rely on discrete acoustic tokens and achieves state-of-the-art performance on the zero-shot sound cloning benchmark Seed-TTS-Eval, supporting timbre fine-tuning and speech editing.
SpaceX AI officially launched the Grok 4.6 model, further enhancing its capabilities for long-duration agent tasks, complex interactions, and visual processing, enabling it to handle complex tasks such as data research and codebase analysis. In the AA Intelligence Index benchmark test, Grok 4.6 achieved the same score as GPT-5.6 Sol.
SpaceX AI has launched Grok Bot, an AI agent product that can log into a user account and run independently in the cloud, performing tasks 24/7. Each bot has its own dedicated cloud computer, supports parallel collaboration among multiple bots, communication between bots, and the ability to create new bots autonomously. The Grok Bot product prototype is the Agent project codenamed "Sand" before xAI acquired Cursor.
DeepSeek has released the official version of DeepSeek V4 Pro, which is now available on its app, web platform, and API. This version features significantly enhanced agent capabilities, demonstrating excellent performance on benchmark datasets such as Terminal Bench and DeepSWE. The API now natively supports the OpenAI Responses API format and is compatible with Codex; three new thinking modes—low, high, and max—are now available.
The Qwen team has open-sourced the Qwen3.8-2.4T-A95B model, marking the first time weights have been made available at the Qwen-Max level. This MoE model has a total of 2.4T parameters, 95B activations per token, and natively supports 256K context, significantly enhancing capabilities for programming, office work, research, and long-cycle agents. It has achieved leading performance in multiple benchmarks, including PaperBench and TerminalBench, and supports deployment with frameworks such as SGLang and vLLM. It features a default thinking mode and supports adjustable inference depth.
Open-world modeling company LTX has released LTX-2.5, natively integrating ComfyUI. On two NVIDIA GB200 GPUs, it generates a 10-second 720p video in just 6.8 seconds, while the API version outputs 1080p in 23.7 seconds. The new model features higher pixel fidelity, multi-camera continuity, and diffusion-fidelity rendering technology. Users can access it via Hugging Face, ComfyUI, and the API.
The official API version of DeepSeek V4 Pro (DeepSeek-V4-Pro-0813) is now available, with the API call name remaining unchanged. The model context length is 1MB, the maximum output is 384KB, and the pricing is 3 yuan per million tokens for input and 6 yuan per million tokens for output. Cache hit rate is as low as 0.025 yuan. Benchmark tests show that the model performs close to Claude Fable 5 in Agent benchmarks such as Terminal Bench, and even surpasses it in some challenging tests.
Microsoft has released an upgraded programming model, MAI-Code-1.1-Flash, with significantly improved coding capabilities: the model performs 22% better in Terminal-Bench 2.1 tests, 15% better in .NET tasks, and adds native visual capabilities. The new model also boasts a 25% faster token generation speed, a 25% lower cost, and a price reduced to one-quarter of the original. It is currently available on Copilot platforms such as VS Code and JetBrains, and the older version will be completely discontinued on September 10, 2026.
Doubao has launched a special student promotion, offering university students 2.5 times the free credit limit and a discounted price for the professional version to assist with complex learning tasks such as research and creative work. Users can complete the verification process by sending "I want to claim student discounts" through the Doubao PC client, or by verifying their Douyin account in the settings interface to obtain the benefits.
Bilibili's open-source industrial-grade zero-shot speech cloning model, IndexTTS-2.5, has only 0.8B parameters and supports cross-language transfer in five languages: Chinese, English, Japanese, Spanish, and Arabic. The model features a redesigned inference architecture, resulting in a 2.28x speedup and a real-time performance rate as low as 0.20. It also supports stepless speech rate adjustment from 0.5x to 2.0x and provides pronunciation intervention at three granularities: Pinyin, CMU phonemes, and Japanese kana.
Manus has announced its separation from Meta and resumption of independent operations. Manus was acquired by Meta for over $2 billion last December, and the two companies restructured due to cross-border regulatory scrutiny. Manus is asking some users to back up their data by August 23rd, with service restored on August 25th. Affected users will not be charged any fees during the transition period and will receive a return bonus.
Tencent has launched "Little Goose Desktop," a desktop customization tool for Windows, targeting the market with its core selling point of being free and ad-free, offering millions of 4K ultra-high-definition dynamic and static wallpapers. The product features a lightweight AI chat window in the upper left corner of the desktop, enabling users to complete to-do lists and search for files. Aimed at students and general individual users, the tool differentiates itself from the office-oriented product WorkBuddy, securing a foothold in high-frequency Windows desktop scenarios through wallpaper customization.
ZCode, a product of Zhipu AI, has undergone a comprehensive upgrade, launching four new features: Goal, Subagents, Remote Control, and Off-Peak Tasks. Goal mode allows users to set acceptable goals, break down complex tasks, execute them, and iteratively deliver them. Subagents enables parallel collaboration among multiple agents. Remote Control supports remote management via mobile phone/WeChat. Off-Peak Tasks can be executed without credits during off-peak hours. Simultaneously, GLM Coding Plan credits have been uniformly reset, with all users' credits fully restored.
Meta has released the Muse Glimmer, a 30 billion parameter model. Designed for local, resident operation, the model utilizes 4-bit quantization and DFlash speculative decoding technology to run smoothly on devices with 24GB of video memory, such as Macs and consumer-grade GPUs. It supports offline operation, multimodal perception, tool calls, and code writing.