OpenAI stated that smaller-scale training and evaluation have continued, but the planned largest-scale frontier reinforcement learning training remains paused, not a complete halt to model training.
Microsoft has open-sourced the Mage family of 4B-parameter multimodal models, including the streaming video/image understanding model Mage-VL and the image generation and editing model Mage-Flow. Mage-VL uses Codec-Native encoding, reducing visual tokens by 75% and speeding up video inference by 3.5 times, supporting real-time streaming understanding. Mage-Flow supports text-to-image processing and instruction editing at any native resolution from 512 to 2048. The Turbo version requires only 4 sampling steps, and a single A100 GPU can output an image in 0.6 seconds.
The pricing page for BigModel by Zhipu lists GLM-5.3 as follows: context length is 1 million tokens, and input, output, and cache hit prices are 8 yuan, 28 yuan, and 2 yuan per million tokens, respectively.
AI video unicorn Higgsfield has raised $400 million in funding, valuing the company at $5.4 billion (approximately 36.6 billion yuan) post-money. Investors include DST Global, Goldman Sachs, Liberty Global, and Intel. The company's annualized revenue has reached $700 million, growing approximately 34 times in one year, and its valuation has increased by 315% in eight months. The platform boasts over 30 million users across 238 countries and regions.
ByteDance has reached a global agreement with the Motion Picture Association of America to strengthen intellectual property protection for its AI video and image generation models.
Greg Brockman, co-founder of OpenAI, suggested that companies should strengthen basic security controls and gradually use AI agents for code review, vulnerability investigation, and alert classification.
Groq announced the completion of a $350 million funding round led by Disruptive, with NVIDIA planning to participate, valuing the company at $3.5 billion; the announcement defines this round as Series A.
OpenAI will provide a total of $1 million in funding and up to $1 million in API credits to 14 independent projects researching topics such as economic opportunities and social resilience.
Doubao's work task mode adds the ability to remotely operate a computer via mobile phone, allowing users to handle temporary work tasks that require computer execution on their mobile devices.
Cumora is a cross-platform team chat tool for collaboration between humans and AI agents, supporting both cloud-based agents and integration with local Claude Code or Codex.
Cursor Origin has begun offering repository hosting, code browsing, pull requests, and GitHub synchronization features, and is gradually rolling out early testing to paid plan users.
GitHub experienced a critical service incident on August 17, affecting APIs, Issues, Pages, Git operations, Webhooks, Actions, Pull Requests, and Copilot.
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.
Anthropic has begun previewing the /design command for Claude Code, allowing users to invoke Claude Design's artboard workflow from the CLI or desktop; Claude Design is currently in the testing phase.
OpenAI will lease approximately 8 GW of IT data center capacity at the Ports-Pike campus in Ohio, with the first 800 MW expected to be operational in 2028; NVIDIA will provide funding and credit support for the initial 4.25 GW portion.
The Financial Times reported that OpenAI has disbanded its independent Preparedness team and reassigned biometrics and cybersecurity responsibilities to existing teams; this is an organizational restructuring and should not be interpreted as a cessation of security work.
Dario Amodei responded to criticisms regarding AI risk communication and regulatory stances, advocating that regulation should focus on cutting-edge, high-risk capabilities, and believing that public trust ultimately needs to be rebuilt through real achievements in fields such as medicine and biology.
Cursor has officially confirmed that eligible SuperGrok Heavy users who link their Grok and Cursor accounts can receive Cursor Ultra for free; as long as the Heavy subscription remains valid, the benefits will continue.
OpenCode Go currently offers usage caps of $12 per 5 hours, $30 per week, and $60 per month, with DeepSeek V4 Flash being one of the supported models.
The DeepSeek API has adopted peak-valley pricing: peak hours are 09:00-12:00 and 14:00-18:00 Beijing time, and the peak price for the same model is approximately twice the off-peak price.
The Codex head confirmed that the product "will have Astra" in the future, but OpenAI has not yet released an official release page, specific capabilities, price, or launch date for Astra.
Codex officials have released a manual configuration method for the GPT-5.6 Sol million-token context, but this is not the default setting; inputs outside the default window will result in higher limits or API costs.
MyContext is a local-first desktop application that attempts to bring together everyday communication and knowledge-based work contexts into a unified environment; the project uses the Elastic License 2.0.
Kunlun Tech has launched Mureka V9.5, an AI music generation model based on the MusiCoT framework. V9.5 represents a significant leap forward in arrangement spacing, vocal realism, and the accuracy of intent, achieving a vocal quality rate of 61.0% and a control quality rate of 97.0%. The model has also made breakthroughs in the Chinese folk music genre, with more accurate pronunciation, more restrained harmonies, and more expressive arrangements. Based on over 25,000 user feedback iterations, Mureka V9.5 aims to move beyond simply generating music to creating tracks worthy of repeated listening.