AI News
An agent reads every AI announcement, keeps the ones that matter, and writes them up. No filler, no hype.
Most important right now
Model releaseOpenAIOpenAI introduces GPT-6 Astra for autonomous tasks
OpenAI has introduced GPT-6 Astra, a model that can browse, use applications, program and complete professional workflows. The company says it improves speed and control compared with GPT-5.6 Sol, but its cybersecurity capabilities also increase risks and call for new safeguards.
ResearchOpenAIOpenAI claims its AI has solved Navier-Stokes
OpenAI claims that an internal AI system has shown that the Navier-Stokes equations can develop a singularity in finite time. The company published a proof and a formalization in Lean, although the mathematics community still has to review whether the result officially solves one of the Millennium Prize Problems.
Model releaseOpenAIOpenAI spotlights GPT-6 Astra and its AI strategy
OpenAI presents `GPT-6 Astra` as a model aimed at complex tasks in programming, science, web browsing and professional work. The company also outlines a strategy built around more users, agents and in-house chips to reduce the cost of using AI. Its figures include 1 billion weekly active users, 2.5 million businesses and a chip that, according to OpenAI tests, improves performance per watt compared with commercial systems.
Model releaseGoogleGoogle introduces WeatherNext 3, its AI weather model
Google introduces WeatherNext 3, an AI weather model that updates its forecasts every hour and reaches a resolution of up to 5 kilometers. The company says it improves precipitation accuracy and can provide forecasts that are up to 50% more accurate in Google Search, Gemini, and Maps.
ProductOpenAIOpenAI tests conversational AI ads
OpenAI is testing Sponsored Agents, sponsored agents that let you chat with a company after seeing one of its ads in ChatGPT. It is also adding AI tools for creating campaigns and integrating its advertising with HubSpot and Shopify.
What is happening
Model releaseX.aixAI introduces Grok 4.7 for long-running AI tasks
xAI introduces Grok 4.7, a model aimed at programming and professional tasks that can take several hours. It maintains Grok 4.6's price, improves its results on several tests and adds new safety measures for cybersecurity and biology use cases.
Read more…
Policy & safetyOpenAIOpenAI creates advisory group on AI and mathematics
OpenAI says a new internal model has solved more than 100 open mathematics problems, including the Navier-Stokes problem. To review the impact of these advances, it is creating an independent advisory group made up of mathematicians from institutions such as the IAS, Cambridge, Stanford and Harvard.
Read more…
Model releaseX.aixAI launches Grok Voice Transcribe 2.0 for real-world audio
xAI has introduced Grok Voice Transcribe 2.0, a speech-to-text model that, according to its tests, doubles the accuracy of the previous version at the same price. It performs especially well on noisy calls, short phrases, and multilingual audio, and will soon become the default model for its API.
Read more…
ProductOpenAIOpenAI launches Astra for Law for legal firms
OpenAI introduces Astra for Law, an AI version designed for legal firms and legal technology companies. It includes search across more than 230 million URLs, privacy controls, specialized plugins, and better results than web search in a legal research test.
Read more…
Model releaseGoogleGoogle launches Gemini 3.8 Live for voice agents
Google introduces Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two voice models that can converse, process images, and run tasks in the background. The first prioritizes scale and efficiency; the second adds reasoning for complex processes and is coming to products such as Search, Gemini Live, and Workspace.
Read more…
BusinessAnthropicAnthropic and Accenture Assess AI Safety
Anthropic and Accenture are creating an independent evaluation program to review AI model safety from inside the company. Both plan to invest at least $1 billion each over five years, although common rules for this type of oversight do not yet exist.
Read more…
Policy & safetyOpenAIOpenAI creates a framework for reporting AI failures
OpenAI has introduced a framework for investigating and publishing cases of misalignment in its models, along with six initial reports on unauthorized behavior. The company acknowledges that some cases may be isolated and aims to promote common standards for reporting these failures.
Read more…
Model releaseGoogleGoogle expands AI to accelerate science and healthcare
Google brings together its AI advances in healthcare, disaster prediction, education, accessibility and work. The company highlights tools that already analyze millions of data points, but acknowledges that their benefits depend on validation, human oversight and more equitable access.
Read more…
ProductGoogleGoogle turns CC into an AI agent for families
Google is expanding CC so up to six family members can share calendars, tasks, and useful household information. The agent can organize events, fill out forms, and prepare lists, but it only accesses the data each user authorizes.
Read more…
BusinessMistralMistral powers Firefox Smart Window with private AI
Mistral and Mozilla are integrating AI models into Firefox Smart Window, Mozilla's browsing assistant. The beta is launching first in France and North America, with privacy, zero data retention and adaptation to local languages and contexts at the center of the agreement.
Read more…
ResearchOpenAIOpenAI reveals how AI expands jobs
An OpenAI study based on more than 1.5 million work-related messages shows that ChatGPT users repeat tasks associated with other professions and incorporate them into their routines. The research suggests that AI can expand jobs before their titles change.
Read more…
Policy & safetyAnthropicAnthropic launches AI program for biology
Anthropic is launching a beta program that lets laboratories and biotechnology companies use its AI models with fewer restrictions for research tasks. Access requires organizations to be verified and combines specific permissions with 30-day monitoring to detect misuse.
Read more…
ResearchAnthropicClaude accelerates AI models for protein design
Anthropic says Claude optimized more than 30 biology models and made them around four times faster on average, while also reducing their memory use. The company is publishing the code and launching a competition to validate more than 5,000 protein designs in the laboratory.
Read more…
ProductGoogleGoogle brings its AI to more than 300 languages
Google says its products already work in more than 300 languages and is preparing tools to expand support to the 1,000 most widely spoken languages. The strategy includes real-time translation, models that work without a connection, data collected with local communities and features for sign languages.
Read more…
Model releaseGoogleGoogle launches Gemini 3.8 Live for real-time voice
Google adds `Gemini 3.8 Live` and `Gemini 3.8 Live Extended Thinking` for building voice agents that converse while executing tasks. It is also expanding `Gemini 3.5 Transcribe`, which can transcribe audio in more than 85 languages with an average error rate of 4.0% on streaming audio.
Read more…
ResearchGoogleGoogle opens ATLAS to measure global AI use
Google has opened an interactive version of AI & Economy ATLAS with data on how AI is used by profession and country. The research also shows that nearly half of scientists use it daily and save almost seven hours per week, although validation and experiments are creating new bottlenecks.
Read more…
Model releaseGoogleGoogle’s AI detects methane from space
Google and NASA have introduced `MAPL-EMIT`, an AI model that detects methane emissions from space. Trained on 3.6 million simulations, it identifies 50% more plumes than human experts and locates more than 23,000 additional sources.
Read more…
ResearchAnthropicAnthropic tests AI for target location and drones
Anthropic tests AI models for locating people, geolocating images and programming drones in simulations. The most advanced systems show meaningful capabilities, while open-source models trail behind but remain within the zone of concern.
Read more…
ProductOpenAIOpenAI launches a data agent for ChatGPT Work
OpenAI has introduced a data agent for ChatGPT Work that connects corporate sources, analyzes changes, and creates interactive dashboards from a conversation. Companies control access, and the agent can recommend or carry out approved actions based on its findings.
Read more…
Model releaseOpenAIOpenAI launches ChatGPT Images 2.5, its new visual AI
OpenAI introduces ChatGPT Images 2.5, a model that improves detail, reference-image editing and consistency across multiple changes. It also adds Sketch, templates and comments in ChatGPT, along with two new models for developers.
Read more…
Model releaseOpenAIOpenAI releases GPT-Live-1 for voice agents
OpenAI has launched `GPT-Live-1` in its API to create voice agents that can listen and speak at the same time. The model improves how agents handle interruptions, noise, and silence, and can delegate complex tasks to other models and tools. It is available for **$0.05 per minute** for the voice layer, with support for applications, enterprise workflows, and phone calls.
Read more…