Claude Code Async /goal Mode, API Billing Warning, TabPFN-3 Model Release Today's Highlights Today's top stories highlight crucial updates and warnings for AI developers. Anthropic's Claude Code introduces an asynchronous 'run until done' mode, while a critical PSA warns about unintended API billing; additionally, the TabPFN-3 foundation model has been released for advanced tabular data prediction. Source: https://reddit.com/r/ClaudeAI/comments/1tatxau/claude_code_just_shipped_a_run_until_done_mode/ Anthropic's Claude Code has received a significant update, introducing an asynchronous 'run until done' mode via the new /goal command. This feature, available in v2.1.139, allows developers to specify a completion condition, such as "all tests pass and the PR is ready," enabling Claude Code to continuously work towards that objective without requiring immediate user intervention. This new 'async' functionality is a major step towards more autonomous and persistent developer agent capabilities within the Claude ecosystem. It streamlines development workflows by allowing AI agents to handle longer, multi-step tasks in the background, ensuring the set goal is achieved with minimal developer oversight, making it ideal for complex coding and debugging sessions. Comment: The /goal command for Claude Code sounds like a game-changer for agentic workflows, moving from reactive prompts to persistent, goal-driven execution. Developers can now offload more complex, iterative tasks directly to Claude Code. Source: https://reddit.com/r/ClaudeAI/comments/1tbaq2d/psa_if_your_project_has_an_anthropic_api_key_in/ A critical Public Service Announcement has been issued to developers using Anthropic's Claude Code, highlighting a potential unintended billing issue. If an ANTHROPIC_API_KEY is present in any .env file within a project, Claude Code will reportedly bypass a developer's existing "Max plan" subscription and instead bill their API account directly for usage. This behavior, described by Anthropic as "intentional functionality," can lead to significant and unexpected charges for users who believe their usage is covered by a flat-rate plan. Developers are urged to meticulously review their .env configurations and API key management practices to prevent accidental API key exposure that could result in substantial, unbudgeted costs, as one user reported losing $187 due to this issue. This incident underscores the importance of thoroughly understanding the intricate interaction between local environment variables, service-specific billing mechanisms, and cloud AI service subscriptions. Comment: This billing behavior is a major gotcha for Claude Code users. Always be explicit about API key management and understand how local .env files interact with cloud service authentication and billing. Source: https://reddit.com/r/MachineLearning/comments/1tb3fh5/tabpfn3_just_released_a_pretrained_tabular/ TabPFN-3, the latest iteration of the pre-trained tabular foundation model, has been released, building upon the original model notably published in Nature. This new version significantly enhances its capabilities, now supporting predictions on tabular data with up to an impressive 1 million rows. TabPFN (Tabular Prior-Data Fitted Network) is highly regarded for its ability to perform predictions in a single forward pass, offering exceptional efficiency and strong performance on small to medium-sized tabular datasets. The release of TabPFN-3 marks a significant advancement for data scientists and developers working extensively with tabular data. It provides a robust, pre-trained model that can be directly applied to various classification and regression tasks without the typical overhead of extensive hyperparameter tuning or complex model architecture design. Its core focus on fast, single-pass prediction makes it particularly appealing for real-time applications and scenarios requiring rapid inference and minimal computational resources. Comment: TabPFN-3 sounds like an excellent tool for quick and effective tabular data analysis. The 1M row capacity and single forward pass capability make it a strong contender for many practical applications, avoiding the usual deep learning setup complexity.