August 2025

The release gave developers another open-weight option for agent tasks; its design keeps the parameters active for each token far below the headline count.

The feature reorganized a tutoring exchange around the student’s own reasoning.

Spending on data centers had become the measure investors used to judge companies’ AI ambitions.

The figures, though unconfirmed, gave an outside measure of how fast paid use of AI was growing.

The plan made adoption across the economy, rather than frontier capability alone, the measure of China’s AI policy.

The project tied European demand for AI computing to the region’s power and construction capacity; its scale was a target for 2026.

The dispute showed how closely the leading labs test their models against one another’s.

This was one stage in a phased timetable; the voluntary Code of Practice offered one route to show compliance.

The episode reopened the question of how much a generator should refuse, and who decides.

The release gave developers OpenAI models they could deploy and customize on their own hardware.

The update continued the labs’ competition over models for long, multistep work such as software development.

DeepMind listed its limits, including restricted actions, imperfect geographic accuracy and difficulty simulating interactions among several agents.

Coding agents that work in the background moved from preview to a general product.

The price removed a procurement barrier for agencies trying the tool.

The design built a choice about how much reasoning to spend on a question into the product itself.

Its system and model evaluations measured different configurations, which matters when comparing scores or reproducing results.

The arrangement turned an export restriction into a revenue share, an unusual approach to trade control.

The model addressed one step in a robotics toolchain: interpreting the physical world before acting in it.

Coming a week after OpenAI’s similar offer, it showed the leading labs competing on price for government customers.

The change made long-context work a practical option on Anthropic’s mid-sized model.

Child-safety advocates called on Meta to publish its revised rules so that parents could see how its chatbots treat children.

The round backed a strategy built on business and government customers rather than consumer chat.

Self-supervised training reduced the need for labeled images, and the same features could serve many specialized applications.

The funding committed support for open models built for researchers; the models themselves were still to be developed.

State regulators began applying consumer-protection law to companion chatbots.

The framework separated changes to an agent’s working context from changes to its model or design, and mapped open research directions.

The review placed adaptation within a continuing cycle of action and feedback, and made evaluation a central question.

The product combined reading, synthesis and content creation in one workspace.

The pause followed months of unusually expensive recruiting for frontier AI talent.

The release brought a reasoning mode and ordinary responses into one model.

Experimental agentic features announced at the same time remained more narrowly available.

Meta, which builds much of its own infrastructure, turned to a competitor’s cloud for additional capacity.

The work linked generated protein sequences with laboratory testing at an early, cell-based stage.

Meta supplemented its own models with technology from an independent image lab.

The letter signaled coordinated attention from state law enforcement to children’s use of chatbots.

The suit carried the competition among AI assistants into antitrust law.

The release made iterative editing a central use of image generation; Google noted remaining limits in text rendering, fine detail and consistency.

The council gave Anthropic a standing source of advice from former officials as it sought government work.

Even record results were judged against very high expectations for AI spending.

The releases showed Microsoft building more of the models behind its products alongside its partnership with OpenAI.