Codex Resets Usage Limits for All Paid Users, Will Reset Again on Monday
要闻
Codex lead Tibo and Claude Code creator Boris Cherny clashed after a user claimed they were banned from Claude Code for using other models. Tibo subsequently reset usage limits for all paid users of ChatGPT Work and Codex. Some users complained that the reset was just a 'performance' since the normal weekly quota had just refreshed, and Tibo responded that it would reset again on (local time) Monday.
Recently, a developer said that after he followed the solution previously shared by Tibo to connect an OpenAI model to Claude Code, his Anthropic account was quickly suspended. He publicly appealed and asked Tibo for help. Codex lead Tibo retweeted this, saying that although he does not work at Anthropic, it would indeed be strange if users were banned just for running other models using Claude Code's shell, and asked if there were similar cases.
Claude Code's lead creator Boris Cherny subsequently responded. He first joked about inviting Tibo to work at Anthropic, then clearly stated that Anthropic does not ban accounts because users use other models in Claude Code, and that this was more likely caused by another account risk-control classifier being triggered; the team is investigating. Boris further explained that Claude Code itself supports connecting other models via tools like LiteLLM, but the tool design and prompt optimization for different models are not fully universal.
After the incident was essentially resolved, Tibo said that 'harness freedom is important, and users should be able to decide which model is best for themselves,' while responding to Boris's poaching attempt by saying he would not leave OpenAI. He then said that to celebrate GPT-5.6 Sol being able to run in Claude Code, he announced a reset of usage quotas for ChatGPT Work and Codex paid users.
Some people mocked this reset as 'a performance,' since the normal weekly quota had just refreshed, and Tibo then responded in kind, saying that (local time) Monday there would be another 'performance reset.'
SpaceXAI Releases Grok Imagine Image 2.0 Image Model
模型发布
SpaceXAI has released its next-generation image generation model, Grok Imagine Image 2.0, now fully available in relevant products as Quality Mode, with API access coming soon. The model focuses on precise editing, adding features such as Magic Wand, Segmentation, Smart Resize, background removal, and multi-reference image editing, along with a series of template workflows. It ranks second in the world in both text-to-image and image editing leaderboards on Arena, behind OpenAI's GPT Image 2.
SpaceXAI officially released its next-generation image generation model, Grok Imagine Image 2.0, which is now fully available at grok.com/imagine and on iOS and Android apps in Quality Mode, with API access coming soon. The model is centered on precise editing, adding features such as Magic Wand, Segmentation, Smart Resize, background removal, and multi-reference image editing, along with template workflows for scenarios like photo editing, product images, avatars, and icons. According to official data, Grok Imagine Image 2.0 ranks second in the world in both the Text-to-Image and Image Edit leaderboards on Arena, trailing only OpenAI's GPT Image 2.
MiniMax Plans to Open-Source Unified Image Model Supporting Text-to-Image and General Editing
模型发布
MiniMax plans to open-source a unified image model that supports Text-to-Image and general image editing. It shares a VAE encoder with H3 and has a dedicated decoder, currently in the post-training optimization phase.
MiniMax H3 technical team revealed in a Reddit AMA this week that they plan to open-source a unified image model that supports both Text-to-Image and general image editing, currently in the post-training optimization phase. The model shares the same base architecture lineage as H3, uses H3's VAE Encoder, and features a separately designed VAE Decoder for image generation. MiniMax envisions a workflow where the image model generates a first frame, which is then fed to H3 for video generation. The team also mentioned that H3 was previously trained only on predicting the last frame from a first frame and text description, with no dedicated image editing training, yet it showed strong zero-shot capabilities in several image editing evaluations, which is an important basis for further expanding the image model.
Apple Official Documentation Confirms Alibaba Qwen Integration for Apple Intelligence on China-Market Macs
产品应用
Apple updated its Mac User Guide, confirming for the first time in official documentation that Apple Intelligence supports the Alibaba Qwen extension. After enabling the Qwen extension, users can get more in-depth answers through Siri and use writing tools for text and image creation. However, the relevant content has now been removed, seemingly because the feature has not yet been officially launched.
Apple updated the Mac Simplified Chinese user manual, adding a chapter titled "Using Qwen with Apple Intelligence on Mac," marking the first time the official documentation confirms that Apple Intelligence integrates Alibaba's Qwen extension. Once users enable the Qwen extension, they can ask Siri for more in-depth answers and use writing tools for text and image creation. Enabling the extension requires logging into a Qwen account, users can control when to use Qwen, and the system will ask for confirmation before sharing attachments such as photos or files. However, the chapter has now been removed, likely because the feature has not yet officially launched.
OpenAI has acquired NextSlide. The company's core product converts documents and more into polished, editable presentations. The founder said the deal actually closed earlier this year, and the team has now joined OpenAI to assist in developing ChatGPT.
OpenAI has acquired presentation startup NextSlide, though the specific transaction amount was not disclosed. NextSlide's core product can convert prompts, notes, documents, or research into polished and editable presentations. NextSlide founder Ahmed Beshry said the actual acquisition took place earlier this year, and the team has now joined OpenAI to help develop ChatGPT, continuing the goal of making visual communication more accessible.
Sources Say Alibaba Plans to Follow Moonshot AI in Collecting Model Commercial Revenue Share
前瞻与传闻
According to reports, Alibaba plans to charge revenue share from large reasoning service providers deploying Qwen3.8-Max, with the specific percentage yet to be determined. The move reportedly follows Moonshot AI's terms for Kimi K3, where any entity offering the model with annual sales exceeding a certain amount must enter a commercial agreement and provide a revenue share. Sources say Moonshot AI's required revenue share can be as high as 30%.
According to Reuters, sources familiar with the matter said that Alibaba plans to implement a new strategy for its next-generation open-source model Qwen3.8-Max next week, requiring revenue sharing from users who derive significant commercial benefits through the model. This move is similar to the terms set by Moonshot AI in its open-source model Kimi K3, where any entity providing services for the model with annual sales exceeding $20 million must enter into a commercial agreement with the company. Sources said that Moonshot AI requests a revenue share of up to 30%, while Alibaba's specific commission ratio is still under discussion and has not yet been clarified.
WeChat Grayscale-Tests "AI Writing Assistant" and "AI Comments" for Moments
前瞻与传闻
According to reports, WeChat is testing "AI Writing Assistant" and "AI Comments" features for Moments on a small scale. The feature, based on AI assistant Xiaowei, can generate copy in multiple styles.
It is reported that Tencent customer service responded to confirm that WeChat is conducting a small-scale gray-scale test of the 'AI Assist Writing' and 'AI Comment' features on Moments. When users click 'AI Assist Writing' on the Moments editing page, it will redirect to the conversation interface of the AI assistant Xiaowei. Xiaowei can recognize the image-text content being edited and generate candidate copy in three styles: 'concise', 'playful', and 'literary'. In the current gray-scale test version, after completing the copy generation and pre-publishing preparation, the final publication may still require manual confirmation by the user.
Note: Content is AI-assisted and may contain hallucinations and errors.
