Stripe is reportedly in talks to acquire OpenRouter, the platform that acts as a gateway for multiple AI models. The deal would value the startup at over $7 billion, consolidating access to models like Anthropic's and others through a single entry point.

Reports indicate OpenAI disbanded its preparedness team tasked with assessing model risks and implementing mitigations. The move has raised concerns about the company's safety priorities.

Anthropic detailed how it will implement invisible watermarks in Claude-generated text to comply with EU AI transparency regulations. The system uses SynthID-Text, technology developed by Google DeepMind that creates detectable patterns in word probabilities.

OpenAI announced funding for 14 independent projects exploring new policy ideas for the AI age. The initiative aims to expand economic opportunities and strengthen societal resilience in the face of technological change.

ChatGPT added Computer History to its macOS desktop app, a feature that converts your actions into training data. The tool learns how you work, suggests automations, and can complete tasks you left unfinished.

Researchers analyzed a full year of production traffic in an LLM serving service, revealing workload evolution patterns and how users interact with multiple models in production. The paper releases a public dataset for future serving optimization research.

A circuit analysis across 46 tasks reveals that LLMs emerge with modular specialization similar to human brain networks: language, formal reasoning, social, and physical reasoning. This finding suggests modularity is a fundamental principle of intelligent systems.

A new paper challenges the common assumption that Language Server Protocol is more token-efficient than lexical search for coding agents. The study finds conditional and often negative results: LSP can cost +6% to +118% more tokens.

SELR is presented, a framework that trains a single model for efficient yet interpretable latent reasoning. The method combines accuracy optimization with the ability to self-decode its own latent representations, balancing efficiency and explainability.

BCMT is proposed, an architecture that avoids the quadratic complexity of self-attention using adaptive causal memory between blocks. The system maintains dense attention only locally and efficiently propagates global context, compatible with standard implementations.

Agentao is introduced, a runtime that governs tool-using LLM agents with a layered architecture separating model-proposed actions from authorized execution. The system mitigates risks of over-privileged actions, prompt injection, and uncontrolled side effects.

Anthropic launched support for configurable system prompts in Claude's API, allowing developers to customize the model's behavior and personality. This feature simplifies Claude's adaptation for specific use cases.

InflationAgent is proposed, a router that measures token inflation (true workflow cost vs theoretical per-token cost) in LLM agent workflows. The system finds inflation up to 4.25x in smaller models and selects models by maximizing accuracy-to-cost ratio.

RubricForge is a method that induces readable evaluation rubrics from labeled data, avoiding the common problem of automatic judges overestimating failed agents. The approach reduces false positives and generates verdicts attributable to explicit criteria.

MBZUAI, Cerebras, and Inception released Jais 2, a family of open-source models optimized for Arabic with 70B and 8B parameter variants. The models achieve competitive performance on Arabic and culturally-specific benchmarks with significantly fewer training tokens.