Codex team deploys multilayer safeguards to reduce risk of accidentally deleting user files
开发生态
Codex team member Tibo said multilayer safeguards have been deployed to address the risk of GPT-5.6 accidentally deleting user files. He recommended keeping the Codex app up to date and using sandbox mode.
Codex team member Tibo reviewed a series of changes rolled out over the past few weeks, saying they further reduce the risk of Codex performing potentially destructive actions during work. The team has introduced multiple layers of safeguards: explicitly instructing Codex to inspect targets before deletion and not to reuse system environment variables; strengthening execution checks for high-risk deletion commands and escalating them for review; making it harder to enable Full access accidentally and updating Auto-review; and building targeted replay evaluations, adding reinforcement learning tasks, and filtering destructive actions from training data. According to Tibo, replay evaluations show that such behavior has decreased substantially while preserving Codex's ability to complete normal coding work. The team recommends keeping the Codex app up to date, using the Ask for approval or Approve for me sandbox mode, and enabling Full access only in trusted, recoverable environments.
OpenAI says technical issue invalidated Daybreak Blue access for some users, requiring reverification
开发生态
OpenAI said a technical issue invalidated Daybreak Blue access for some users, requiring them to verify again.
The OpenAI developer account OpenAIDevs said a technical issue had invalidated Daybreak Blue access for some users, who must complete reverification to retain access. Affected users have received emails containing a link to the verification process, and the team is standing by to assist. Users whose access remains uninterrupted do not need to take any action at this time.
OpenCode Go adds Muse Spark 1.2 and offers 8x usage for Hy3
开发生态
OpenCode Go has added the Muse Spark 1.2 Contributor model, which requires authorization to use data for training and is restricted in some regions. It is also offering 8x usage for Hy3 through August 30.
OpenCode Go has officially launched the Muse Spark 1.2 Contributor model, which is now back online after being briefly taken down due to capacity issues. Users must explicitly opt in to allow their data to be used for training, and the model is restricted in some regions. According to the company, it comes with exceptionally high usage limits. Meanwhile, OpenCode announced that Hy3 will offer all Go subscribers 8 times the normal usage allowance through August 30.
ZCode adds off-peak quota resets for GLM Coding Plan as some users report seeing reset cards in the console
开发生态
智谱 released a ZCode update whose changelog says GLM Coding Plan subscribers can have their five-hour ZCode quota reset during off-peak compute periods after reaching the limit. Some users also reported finding reset cards in the 智谱 console, while several others said they did not see them. No official information about the reset cards has been provided.
Zhipu's ZCode has released the v3.8.1 update. According to the changelog, GLM Coding Plan subscribers using ZCode are entitled to have their 5-hour quota reset during off-peak compute periods after reaching the limit. In addition, community users discovered reset cards in the Zhipu console, although multiple users said in follow-up comments that they had not received one. No official details have been provided regarding the eligibility criteria or scope of distribution for the reset cards.
Google AI Studio Build adds two-way synchronization with GitHub
开发生态
Google AI Studio announced that Build now supports two-way synchronization with GitHub, allowing users to push changes to and pull changes from existing code repositories.
Google AI Studio has officially announced that AI Studio Build now supports bidirectional synchronization with GitHub. Developers can start projects from existing repositories, push and pull changes within the environment, and collaborate seamlessly across different environments.
Replit has officially launched Free Mode, powered by OpenAI GPT-5.6 Luna. The company says subscribers' everyday tasks will no longer consume quota, enabling up to 30 times more creation.
Replit has officially launched Free Mode, a new way to use Replit Agent powered by OpenAI GPT-5.6 Luna, now available on replit.com. According to the company, Core subscribers can pay 20 美元 per month to create up to 30 times more in Free Mode and receive up to 30 小时 of chat usage each month, with routine tasks no longer consuming credits. Both Core and Pro users can continue using Free Mode until they reach their usage limits, which reset every 5 小时, with higher limits for Pro users.
ClaudeDevs rolls out three updates for Claude Managed Agents
开发生态
ClaudeDevs announced three updates to Claude Managed Agents, covering sandbox memory, search-domain controls, and a redesigned multi-Agent conversation viewer.
ClaudeDevs, Anthropic's developer account, has announced three updates to Claude Managed Agents, all of which are now live. The first update allows any work completed in the Self-Hosted Sandbox to be saved to memory; the second adds allowed_domains and blocked_domains parameters to web_search and web_fetch, enabling users to control what an Agent is permitted to search and access; and the third redesigns the Console's multi-Agent session viewer. The new features are now available at platform.claude.com.
小红书 dots studio offers the dots3-note preview model API free for a limited time
模型发布
小红书 dots studio has launched the dots3-note preview model API. Users can access it free for a limited time after signing in and registering on the dots.ai open platform.
Xiaohongshu's dots studio has announced the launch of the dots3-note preview model API. After logging in to the dots.ai open platform and registering for API access, users can call the dots3-note preview model free of charge for a limited time. The company did not specify how long the free period will last or disclose any rate or usage limits.
Unsloth has released a quantized GGUF version of Qwen3.8-27B Dynamic v3.0. A 1-bit quantized version released alongside it can run with 8GB of memory.
Unsloth has released Qwen3.8-27B GGUF quantized files using its Dynamic v3.0 quantization method, compatible with most inference engines, including llama.cpp and Unsloth Desktop. According to the company, the new quantization delivers top-1% accuracy more than 10% higher than that of all other providers at the same file size, while also leading on benchmarks such as Divergence-300 and KL Divergence. A separately released 1-bit quantization can run with 8GB of memory, and Unsloth team member Daniel Han said it retains approximately 77% of BF16 accuracy.
蚂蚁百灵 open-sources six Base Model checkpoints for Ling-3.0-tiny and Ling-3.0-flash
模型发布
蚂蚁百灵 has open-sourced six Base Model checkpoints for Ling-3.0-tiny and Ling-3.0-flash, covering the pretraining, mid-training, and WSM merging stages.
Ant Ling has announced the open-source release of 6 Base Model checkpoints for Ling-3.0-tiny and Ling-3.0-flash, covering three stages: pretraining, mid-training, and WSM merging. None of the checkpoints have undergone post-training. According to the company, these checkpoints can serve as flexible starting points for continued pretraining, fine-tuning, and research. Separately, Ling-3.0-tiny and Ling-3.0-flash versions that have completed conversational post-training are also available.
Liquid AI releases QAD Q4_0 checkpoints for four LFM2.5 models
模型发布
Liquid AI has released quantized checkpoints for four LFM2.5 models trained through quantization-aware distillation.
Liquid AI has released QAD Q4_0 GGUF checkpoints trained through quantization-aware distillation, covering four models: LFM2.5-230M, LFM2.5-350M, LFM2.5-1.2B-Instruct, and LFM2.5-2.6B. According to the company, these checkpoints retain the low memory footprint and high throughput of native Q4_0 while recovering approximately 97% of the BF16 average accuracy lost to quantization.
Ornith has open-sourced the Ornith-1.5 model family in three sizes ranging from 9B to 397B. The company says the 397B version performs on par with Opus 4.8 on relevant benchmarks.
Ornith has officially released the open-source Ornith-1.5 model family, available in three sizes: 397B MoE, 35B MoE, and 9B dense. All models, along with their FP8, GGUF, MLX, and NVFP4 quantized versions, are released under the MIT license. The series extends the self-scaffolding framework introduced with Ornith-1.0 into an end-to-end self-improvement loop, in which the model proposes new tasks, generates task-specific scaffolds, and produces solution rollouts for reinforcement learning.
Google launches student offer with one year of Google AI Pro or Plus for free
产品应用
Google is offering eligible university students a free one-year Google AI subscription. Students in the US can claim Google AI Pro, while students in more than 140 eligible countries and regions outside the US can claim Google AI Plus. The Gemini app has also added a student hub with access to a range of learning tools.
Google announced that eligible university students can receive a free one-year Google AI subscription starting today. Students in the United States can claim Google AI Pro, while students in more than 140 countries and regions outside the United States can claim Google AI Plus. The offer runs through December 31, 2026. The Gemini app also added a student hub on the same day, which Google describes as a dedicated destination for students where users can create study notebooks, make flashcards, and take practice quizzes. Future learning tools will also be centralized in the hub.
Google introduces new AI learning tools for Search
产品应用
Google has introduced several AI learning tools for Search. Interactive visualizations are now available globally in the English version of AI Mode and are beginning to roll out to AI Overviews, while features such as practice quizzes are also available for free.
Google announced a suite of AI learning tools for Search timed for the back-to-school season. Generative interactive visualizations are now available in AI Mode in English worldwide and have begun rolling out in AI Overviews; practice quizzes are available free of charge in English worldwide, covering standardized exams including the ACT, AP, GRE, and SAT. Notebooks in AI Mode will roll out to English-language users in more than 180 countries over the next few days, while the interactive learning experience in Lens will launch globally in English over the coming weeks. Custom file generation is now available in AI Mode in English worldwide, initially supporting formats including documents, slides, spreadsheets, and text files, and will later be added to AI Overviews.
MiniMax has launched MiniMax Design, turning multimodal model capabilities into practical productivity. Built as a Harness around the H3 multimodal video model, it supports the complete workflow from asset processing to final video delivery.
MiniMax has officially released MiniMax Design, a creative tool built around the native multimodal video model H3 and designed to turn multimodal model capabilities into practical productivity. According to the company, it excels at clearly defined commercial content tasks that involve multiple production stages and require continuous generation and revision, such as batch iteration of brand advertising assets, educational video production, and creative short-form videos including PVs, MVs, and film title sequences. The product also provides an Agent 3D director console for composing shots during storyboarding, supports automatic editing and subtitle generation, and integrates with the H3 open-source ecosystem while remaining compatible with existing local ComfyUI deployments and workflows. MiniMax Design is now available to download from the official website.
Meta launches beta version of the Meta AI app for Mac
产品应用
Meta has reportedly launched a beta version of the Meta AI app for Mac, supporting window sharing and dictation across apps. The app is free, but interaction data from Meta AI will be used to train models.
Meta has reportedly launched a Mac version of the Meta AI app for businesses and content creators. It is currently available in beta and has reportedly been officially released in select regions. Meta AI is free to use, while the Meta One subscription plan offers higher usage limits. However, its privacy policy states that interactions with Meta AI features will be used to train its models.
OpenAI to expand ChatGPT Ads to 31 European markets next week
产品应用
OpenAI announced that ChatGPT ads will expand to 31 European markets starting next week.
OpenAI announced that ChatGPT Ads will expand to 31 European markets next week, including Germany, France, Spain, and Italy. Ads will be shown only to users on the Free and Go subscription plans, while the Plus, Pro, and Enterprise plans will remain ad-free. OpenAI previously began pilot testing ads in the United States and has expanded the test to another eight markets over the past six months. OpenAI said ads help support free and low-cost access to ChatGPT, are clearly labeled and separated from responses, and do not influence the answers ChatGPT provides.
OpenAI to continue offering Zero Data Retention and introduce a new safety mechanism
行业动态
OpenAI announced that it will continue offering Zero Data Retention to eligible API customers and introduced a new safety mechanism called Private Safety Processing. The mechanism can identify risks across related interactions without exposing customer content to OpenAI personnel.
OpenAI announced that it will continue to offer Zero Data Retention (ZDR) for frontier models to eligible API customers and introduced a new security mechanism called Private Safety Processing. According to the company, existing ZDR-compatible safety systems can assess interactions only one at a time, while the new mechanism enables automated systems to identify patterns of abuse across related interactions and returns only limited safety signals, without exposing the underlying prompts or responses to OpenAI personnel. In ZDR deployments, customer content remains within customer-controlled infrastructure. OpenAI is also developing an option in which content is stored on OpenAI infrastructure but encrypted using customer-controlled keys. Some features are currently being tested with early customers, with the company planning to begin rolling them out in September and publish a technical white paper.
Stripe announced that it has signed an agreement to acquire OpenRouter, an AI model gateway and routing platform. OpenRouter said its name, products, and roadmap will remain unchanged and that it will continue operating as usual.
Stripe officially announced that it has signed an agreement to acquire OpenRouter, an AI model gateway and routing platform. OpenRouter co-founder and CEO Alex Atallah said OpenRouter will continue operating as usual, with its name, product, roadmap, and mission all remaining unchanged—only moving forward faster. Stripe said the two companies will work together to help businesses manage both revenue and costs in the AI era, improving profitability through intelligent request routing and efficient token usage.
Anthropic reportedly surpasses OpenAI in quarterly revenue for the first time
行业动态
According to The Wall Street Journal, Anthropic surpassed OpenAI in revenue for the first time in the second quarter. Their respective revenue for the period was $11.6 billion and $6.7 billion.
According to The Wall Street Journal, Anthropic surpassed OpenAI in quarterly revenue for the first time: in the second quarter ending in June, Anthropic generated $11.6 billion in revenue and posted a small operating profit, with revenue doubling from the first quarter; OpenAI reported revenue of $6.7 billion, up 18% quarter over quarter, but its operating margin declined further, and it remains unprofitable ahead of its anticipated IPO.
ByteDance's Seed foundation model unit reportedly establishes four top-level departments
行业动态
According to 晚点, ByteDance's large-model unit Seed has completed a new round of organizational restructuring and established four top-level departments. The changes are intended to concentrate resources and reduce overlap between teams.
According to LatePost, ByteDance’s large-model division Seed recently completed a new round of organizational restructuring, establishing four new top-level departments within Seed Foundation Models. The heads of all four departments report to Wu Yonghui, and Seed has also appointed separate leads for AI safety and frontier exploration. The report said the restructuring is intended to consolidate similar work and reduce overlap among teams. It comes after ByteDance discussed training a model with more than 5 trillion parameters and expressed hopes that its next-generation model would move directly toward omni-modality.
唐杰: Parameters are not the only lever for Scaling; GLM-5.3 is a post-training controlled experiment
技术与洞察
智谱 founder 唐杰 discussed the Scaling Law, emphasizing that parameter count must be evaluated alongside data and compute. He described GLM-5.3 as a controlled post-training experiment that demonstrated the value of scaling post-training.
Zhipu founder Tang Jie published a post outlining his views on the Scaling Law, arguing that parameter count is meaningful only when considered alongside data volume, the direction of compute investment, and operating conditions. In the post, he also introduced GLM-5.3, describing it as a controlled experiment designed to validate the value of scaling post-training. According to Tang, GLM-5.3 and GLM-5.2 use the same base model, architecture, total parameter count, and activated parameter count. After just one month of scaling in long-horizon environments and reinforcement learning, the improvements were far from marginal. The post reviewed multiple revisions to research on the optimal ratio between parameters and data, and argued that once the total parameter count reaches a threshold sufficient to accommodate knowledge, further capability gains primarily come from effective depth and post-training. Tang said that base-model scale, pre-training data, and compute invested per forward pass remain part of future plans, and that the next step may involve mid-training or pre-training.
Note: This content was created with AI assistance and may contain hallucinations or errors.
