Moonshot AI (Kimi) launched the new generation model Kimi K3 on July 16. The model has a total parameter count of 2.8T (2800B), supports a 1 million context window, and is released under an open-source license. According to Kimi’s official report, K3’s scores are second only to Anthropic Fable 5 and OpenAI GPT-5.6 Sol, surpassing Claude Opus. Multiple bloggers gave positive reviews after testing: one blogger in tests like “Lifting the World Cup” believed K3’s performance was only second to Fable 5; another blogger claimed that K3 often approached Fable levels in multiple tests and consistently outperformed GPT-5.6. K3’s reasoning chain uses English output. In terms of pricing, the $49 tier cannot use K3, the $99 tier has only 256K context, and at least the $199 tier is needed to experience the full 1 million context. Some bloggers praised it as “the number one domestic model”.<\/p>\n
Sources:<\/p>\n
xAI officially announced that Grok Build is fully open-sourced, while resetting usage quotas for all users. The GitHub repository has received thousands of stars. After open-sourcing, multiple developers analyzed the source code and concluded that the previously rumored “bundling and uploading user codebases” issue no longer exists. One blogger had AI dissect Grok Build’s source code and generated a complete analysis document; another blogger wrote that the repository is an 800,000+ line project, including complete prompts, self-developed and ported tools, memory dreaming, anti-deadloop, and other full components. Musk himself retweeted to confirm this.<\/p>\n
Sources:<\/p>\n
Anthropic and OpenAI announced the reset of AI coding tool usage quotas on the same day. Claude Devs first reset the 5-hour and weekly usage limits for all users; just minutes later, OpenAI announced that ChatGPT’s weekly active users reached 9 million and also reset the quotas for Codex and ChatGPT Work. Multiple bloggers attributed this phenomenon to user benefits from competition. However, some bloggers\u53cd\u9988 that after the reset, Codex’s response speed\u660e\u663e\u53d8\u6162, with task execution times several times longer than before. Tibo(@thsottiaux) stated that the 5-hour limit for Codex Plus\/Pro users has been removed, and they are\u8c03\u7814 user feedback on the unlimited mode.<\/p>\n
Sources:<\/p>\n
Anthropic, together with Blackstone and Hellman & Friedman, jointly initiated Ode, an AI deployment service company focused on result delivery. Ode’s positioning is to provide cutting-edge AI delivery teams for mid-sized enterprises, with the team and enterprise jointly defining high-priority AI projects, using Claude to deploy solutions into operational systems, covering industries like finance, healthcare, retail, manufacturing, software. The investor lineup also includes top institutions like Goldman Sachs, General Atlantic, Apollo, GIC, Sequoia. Ode’s foundation is Fractional AI, acquired in May 2026, with founders Chris Taylor and Eddie Siegel serving as CEO and CTO respectively. Some bloggers commented that model companies entering services upgrades from selling tokens to selling transformation results, posing a direct threat to traditional consulting firms.<\/p>\n
Sources:<\/p>\n
Thinking Machines, founded by former OpenAI CTO Mira Murati, launched its first model Inkling after a year and a half. The model uses MoE architecture, with total parameters 975B and activated parameters 41B, trained on 45T data, supporting text, image, and audio inputs. The model is released under Apache 2.0 open-source license,\u9644\u5e26 a Small preview version with 12B activated parameters. Benchmarks place it in the upper-middle tier of open-source models: HLE with tools 46.0%, SWEBench Verified 77.6%, GPQA 87.2%, comparable to Kimi K2.6 and DeepSeek V4 Pro, but still behind closed-source top-tier. Its training data (post-training part) comes from distillation of multiple open-source models. The model supports fine-tuning and customization on the Tinker platform.<\/p>\n
Sources:<\/p>\n
OpenAI Devs released the company’s first hardware product, the kbd-1.0-codex-micro, a custom mechanical keyboard in collaboration with Work Louder. The keyboard supports custom key mapping,\u9644\u5e26 32 Codex-themed keycaps with different icons, with each Agent key lighting up in different RGB effects based on Codex status. Some bloggers praised the design but found the $230 price high, suggesting users without aesthetic needs could buy cheaper customizable numpads. Other bloggers said they would buy if the keyboard supports Claude Code. The keyboard sold out quickly after release, and OpenAI Devs subsequently added more stock.<\/p>\n
Sources:<\/p>\n
OpenAI announced GPT-Red, an internal automated red team system to discover prompt injection vulnerabilities at scale. GPT-Red is trained to attack GPT-5.6 Sol, helping build stronger defenses. OpenAI stated that by replaying the strongest attacks generated by GPT-Red (never seen during model training), GPT-5.6 Sol proved to be the most robust model against prompt injection to date, with a failure rate 6 times lower than the best model four months ago. Additionally, OpenAI demonstrated improvements in GPT-Live’s multi-task capabilities, with the model able to handle flight queries, weather retrieval, and itinerary planning in real-time during continuous conversations.<\/p>\n
Sources:<\/p>\n
Anthropic’s national security policy director Tarun Chhabra at the Aspen Security Forum commented on US-China AI competition: US AI models lead China by 6 to 9 months, with\u4f18\u52bf mainly in hardware productivity (30 to 40 times in data centers); GLM 5.2 is currently China’s most advanced model; Anthropic\u9996\u6b21\u516c\u5f00\u6307\u8d23 GLM\u84b8\u998f\u4e86 Claude and GPT data; without GPU hardware export controls, Chinese AI should have caught up or even led.<\/p>\n
Sources:<\/p>\n