The Trump administration has partially authorized access to Anthropic's Claude Fable 5 and Mythos 5 models after a two-week national security block. The lift provides strategic relief for Anthropic as it prepares for an initial public offering. The original block also impacted OpenAI, Anthropic's main competitor in the AI race. The move highlights how national security concerns now act as a primary gatekeeper in advanced AI deployment.
Alibaba has banned its employees from using Anthropic's Claude Code at work, citing concerns over alleged backdoor risks. The company is requiring employees to switch to its own coding platform, Qoder. The ban follows a recent accusation by Anthropic that Alibaba had distilled its model, and the discovery of code in Claude Code that checks if the user is from China. Despite Anthropic's access restrictions for Chinese users, Claude Code had remained popular among Chinese programmers.
A user on X expressed amazement at Claude Fable 5, stating that people are constantly building games and some are making money with it. The post promised 10 wild examples, but no details were provided in the tweet text.
Tokenwork.app, a GPT API platform, has launched its fourth public beta, claiming support for GPT-5.5, GPT-5.4, and Codex. New users receive a $1 credit after completing WeChat verification, and referrers earn $1 plus a 5% commission on the invited user's top-ups. Users who comment on the promotional post with a base64-encoded email are offered an extra $5 credit. The operator states the platform is built primarily for their own long-term use, with testing focused on stability, Codex compatibility, API response speed, peak-hour availability, and billing accuracy rather than low pricing.
A v2ex user listed multiple closed-source coding agent products (Tencent's Qclaw Workbuddy, ByteDance's Trae Work and 扣子编程, OpenAI's Codex, Anthropic's Claude) and asked whether open-source alternatives exist. They want to study how such agents internally call MCP skills and encapsulate Python and Node.js environments. The post is a request for learning resources and discussion.
The paper proposes FlashMorph, a method to convert standard Transformers into hybrid attention models by formulating layer selection as a budget-constrained optimization problem. It uses morphable models and linearization regularization to decide which layers keep full attention and which switch to linear attention, considering global interdependencies. This approach outperforms heuristic selection strategies, discovering efficient configurations that maintain strong long-context recall and overall performance. FlashMorph also reduces the computational cost of layer selection itself, making it scalable.