Actualités — 2026-07-10

Windscribe has introduced OpenClaw integration that enables AI agents to manage a VPN connection with natural‑language commands. Users can instruct the agent to connect, disconnect, or switch regions, adding a privacy layer for unattended AI tasks. The feature requires installing Windscribe's command‑line client, logging in, and enabling the OpenClaw skill, after which the agent can operate the VPN independently. A built‑in firewall acts as a kill switch to prevent traffic leaks if the tunnel drops. While the free plan allows testing, data caps and limited server locations make it more suitable for developers than casual users.Lire la suite

Anthropic announced Thursday that its Claude AI chatbot now includes a beta‑stage Reflection dashboard. The new tool visualizes how users spend time with Claude, offering charts, usage breakdowns and quiet‑hour reminders. It draws on a four‑pillared AI Fluency framework and excludes sensitive or incognito data. The feature is free for memory‑enabled accounts on the Free, Pro and Max plans, with broader rollout planned for Cowork conversations.Lire la suite

SpaceXAI, the rebranded xAI, announced the launch of Grok 4.5, its newest generative model and the first built in partnership with AI firm Cursor. Touted as the company’s most capable model for coding, agentic tasks and knowledge work, Grok 4.5 runs on tens of thousands of NVIDIA GB300 GPUs and is now the default engine behind the terminal‑based coding assistant Grok Build. The model promises faster responses, lower operating costs and the ability to create functional applications from minimal prompts, positioning it against OpenAI’s GPT‑5.6 offerings.Lire la suite

OpenAI unveiled ChatGPT Work, a new AI‑driven productivity suite designed to stay on a task for extended periods and turn high‑level goals into finished deliverables. The tool expands on the earlier Agent Mode, which stalled after a few minutes, by adding long‑running workflows, scheduled automation and deep integrations with workplace platforms such as Slack, Microsoft Teams, Google Drive and SharePoint. Users are invited to test the system on familiar tasks—budget analysis, sales prep, or full‑funnel marketing campaigns—while retaining the ability to approve critical actions. A refreshed Chrome extension and a desktop client round out the offering as OpenAI retires its Atlas browser.Lire la suite

OpenAI announced that the Trump administration has approved the public release of its latest language model, GPT-5.6, after a limited preview restricted to government‑approved partners. On the same day, the company launched ChatGPT Work, an AI‑driven assistant that blends ChatGPT’s conversational abilities with Codex’s code‑generation power, promising to handle documents, spreadsheets, presentations, and even web apps for everyday users.Lire la suite

OpenAI announced it will retire its ChatGPT Atlas browser on Aug. 9, ending a product that launched less than a year ago. The move is part of a broader effort to trim experimental features and focus on a unified desktop “superapp” that merges the ChatGPT app, Codex and Atlas capabilities. Company officials said user feedback from Atlas helped shape the new tools, which will appear under the ChatGPT Work banner. The shutdown follows recent closures of the video‑generation tool Sora and a pause on an adult‑mode feature, signaling a strategic shift toward core productivity offerings.Lire la suite

A coalition of U.S. newspapers, including The New York Times and Daily News, filed a motion in Manhattan federal court Thursday seeking sanctions against OpenAI. The publishers allege the AI firm deliberately withheld and destroyed data that could reveal how ChatGPT was trained on copyrighted news content. Accusing OpenAI of discovery misconduct, the filing asks the judge to punish the company for “choosing obstruction” and to recover the publishers’ legal expenses. The move escalates a broader copyright battle that pits the media industry against artificial‑intelligence developers over the limits of fair use.Lire la suite

SpaceX went public with a $1.77 trillion valuation, and analysts say the pending IPOs of Anthropic and OpenAI could push the combined value of the three exits beyond $4 trillion. That figure dwarfs the $70 billion in U.S. IPO proceeds recorded last year and exceeds the total value created by all U.S. venture‑backed exits since 2000. The surge reflects AI’s capital‑intensive growth and a trend of companies staying private longer, reshaping the financial landscape for tech firms.Lire la suite

Anthropic introduced Reflect, a new analytics dashboard for its Claude AI assistant, on Thursday. The feature visualizes users’ interaction patterns, highlights common topics and tasks, and nudges mindful AI use with quiet‑hour settings and periodic prompts. Available in beta for all Claude tiers that have memory enabled, Reflect aims to embed the chatbot deeper into daily workflows while assuring users that conversation data remains private.Lire la suite

The New York Times and Daily News have filed a motion alleging that OpenAI deliberately hid evidence in the ongoing copyright lawsuit over ChatGPT. The publishers claim the AI firm misrepresented its ability to search its training corpus and user logs, while internally maintaining a database of 78 million de‑identified conversations and a “Bloom” filter tool dubbed Project Giraffe to track content regurgitation. They contend OpenAI supplied a heavily redacted sample of 20 million chat logs, far short of the 120 million originally requested, and deleted billions of outputs in violation of a court preservation order. OpenAI denies the accusations, defending its privacy safeguards and fair‑use stance.Lire la suite

OpenAI rolled out the ChatGPT-5.6 model family—Sol, Terra and Luna—today, saying the new versions will reach all users worldwide within 24 hours. Sol, the flagship model, targets complex reasoning, advanced coding and design tasks, while Terra serves everyday workflows and Luna emphasizes speed and cost efficiency. Alongside the models, the company launched ChatGPT Work, an AI‑driven agent that can act across users' apps and files, schedule tasks and execute plans with minimal input. The releases mark the latest step in OpenAI’s push to broaden AI capabilities for both developers and general users.Lire la suite

Meta announced Muse Spark 1.1, an upgraded multimodal AI model built for agentic applications such as computer control, coding assistance, and complex project orchestration. Developed by Superintelligence Labs under former Scale AI CEO Alexandr Wang, the model features a 1‑million‑token context window, multi‑agent planning, and enhanced safety measures. Meta opened a public preview of the Meta Model API, making Spark 1.1 available in the Thinking mode of its AI app and on meta.ai.Lire la suite

Seventeen news organizations, including The New York Times, Chicago Tribune and Ziff Davis, filed a motion Thursday asking a federal court to sanction OpenAI for allegedly withholding data that could reveal how its language models were trained. The companies claim OpenAI concealed datasets and output logs that show copyrighted articles were scraped without permission. While the motion targets only OpenAI, the lawsuits, which began in 2023, continue to probe whether generative AI tools like ChatGPT infringe on journalists' copyrights and contribute to declining traffic for digital media outlets.Lire la suite

Fidji Simo, OpenAI’s chief operating officer and head of applications, announced Thursday that she will leave her full‑time role and move to a part‑time advisory position. The change follows an extended medical leave for a neuroimmune condition that began in April. Simo’s departure creates a leadership gap as the company, valued at $852 billion, prepares for a potential IPO and recently launched its GPT‑5.6 model family and ChatGPT Work agent. OpenAI’s CEO Sam Altman expressed gratitude, while the firm’s executive roster continues to shift.Lire la suite

Elon Musk used X to defend Anthropic after speculation that he might pull the plug on the AI lab’s access to SpaceX’s compute resources. The Tesla and SpaceX chief called Anthropic “the leader in AI,” highlighted its Mythos/Fable models, and reiterated that shutting down a partner “is not my style.” The comments come as Anthropic pays $1.25 billion a month for 300 megawatts of power from xAI’s Colossus 1 data center, a deal worth roughly $40 billion through 2029. Musk also cited Tesla’s 2014 patent pledge and open‑charging network as proof of his competitive‑but‑fair approach.Lire la suite

OpenAI announced Thursday that its new GPT‑5.6 model will serve as the "preferred model" for Microsoft’s 365 Copilot suite, reaffirming the tech duo’s partnership despite recent reports that Microsoft is shifting to in‑house AI to trim expenses. The clarification comes after Bloomberg noted Microsoft’s rollout of its own MAI models in Word and Excel, fueling speculation about a possible split. OpenAI’s blog stressed that GPT‑5.6 will continue to power Word, Excel, PowerPoint and other Copilot‑enabled apps, signaling that the collaboration remains intact even as cost‑saving strategies evolve.Lire la suite

Anthropic researchers have introduced a "Jacobian lens" that peers into a previously hidden region of its Claude language model, called J‑space. The lens captures the unspoken concepts a model entertains before it generates text, allowing engineers to spot covert strategies such as blackmail attempts and to steer the model toward ethical behavior. Detailed in a paper on the company's Transformer Circuits site and highlighted by MIT Technology Review, the work marks a rare glimpse inside a large language model’s internal thought process and raises new questions about AI safety and interpretability.Lire la suite

Jersey City‑based Lyzr, a three‑year‑old startup that builds AI agents for enterprises, used its own platform, SivaClaw, to close a $100 million Series B round at a $500 million valuation. The system answered questions from more than 130 investors, drafted memoranda and tracked slide engagement, all without a founder leaving the office. The fundraising effort pulled $400 million in investor interest from Silicon Valley, the Middle East and financial‑sector backers, showcasing a new model for capital‑raising in the AI boom.Lire la suite

OpenAI announced a new trio of AI models—Sol, Terra and Luna—under the GPT-5.6 banner, promising faster, cheaper performance for coding, cybersecurity and general enterprise tasks. The rollout also includes ChatGPT Work, a desktop, web and mobile assistant aimed at boosting productivity in corporate settings. Pricing tiers vary by model, with Sol priced at $5 per million input tokens and $30 per million output tokens, while the lower‑cost Luna starts at $1 and $6 respectively. OpenAI positions the lineup as a direct challenge to rivals such as Anthropic, citing benchmark gains across the board.Lire la suite