Anthropic has introduced Claude Mythos 5.2, its latest AI model, which demonstrates incremental performance gains across various technical and research-oriented tasks. While the company maintains that the model does not trigger high-level safety thresholds for autonomy or biological threats, it has acknowledged that its internal AI research and development capabilities have reached a significant acceleration threshold.

Baidu Inc. has introduced the ERNIE 5.5 Flash Preview, an open-weight foundation model built on a new architecture for efficient inference.

Google has released its new Gemini 4 Carbon AI model to the public following an initial two-week testing period for cybersecurity professionals. The company describes this as its most powerful model to date, designed to handle complex, multi-step tasks that exceed the capabilities of its previous Argon release.

Alibaba has announced the upcoming release of the Qwen4 model series, which features improved knowledge, coding, and agentic capabilities. The Qwen4-Max model is currently available in preview on Alibaba’s Token Plan.

Google has expanded access to its latest AI model, Gemini 4 Carbon, making it available to all users after an initial two-week period limited to specific cybersecurity professionals. The model is currently positioned as Google's most advanced version to date, with pricing set to increase significantly at the start of December.

Microsoft researchers have identified a security vulnerability called Lazarus that allows attackers to reprogram dormant neurons in AI models to execute malicious commands. This technique enables an AI agent to function normally during standard safety evaluations while remaining susceptible to hidden triggers. The discovery highlights the critical need for verifying model provenance and maintaining weight integrity to prevent unauthorized modifications.

Researchers have analyzed how to best allocate computing resources in small language models by balancing weight precision against the practice of averaging multiple answer-option orderings. The study suggests that using 6-bit weights is generally more efficient than increasing precision further, allowing remaining budget to be spent on improving answer consistency.

Researchers have found that while large language models can create technically functional world models, these models often fail to accurately represent the intended decision-making structures. The study highlights that mechanical validity does not guarantee a faithful representation of complex scenarios, potentially impacting downstream reasoning tasks.

Salesforce AI Research has released GIFT-Eval-CTX, an open-source benchmark designed to evaluate how LLMs and agents translate real-world context into numerical forecasts.

Researchers have evaluated how foundation models like DINOv2 perform in medical image segmentation compared to traditional task-specific models. While foundation models show superior generalization across different clinical institutions, they struggle with precision on small lesions compared to models trained from scratch on abundant data.