AI movements to track before they become full articles.
Chris Fall has resigned three months after taking over the Center for AI Standards and Innovation, which tests unreleased models from OpenAI, Anthropic, DeepMind, Microsoft and xAI. Arvind Raman will serve as acting director. The policy stack is still throwing version conflicts.
Moonshot AI temporarily paused new subscriptions after demand for Kimi K3, its 2.8-trillion-parameter model, strained capacity, Reuters reports. The Chinese startup is also seeking fresh funding before a possible Hong Kong listing. Even 2.8 trillion parameters still need somewhere to sit.
Anthropic is committing C$10 million to Canadian AI research, partnering with Amii, Mila, Vector, CHEO, CAMH and universities. Claude credits will support work in health, safety, low-resource languages, robotics and multi-agent systems. Canada helped light the deep-learning fuse; Anthropic is sending a thank-you note with credits.
Moonshot has released Kimi K3, billed as its most capable model and the first open model near the 3-trillion-parameter scale. Available in Kimi Code, it adds native vision, up to a 1M-token context window and long-horizon coding strengths. Company benchmarks need caution, but the signal is big.
A US official says Nvidia has begun shipping H200 AI chips to China, while Reuters reports that ZTE is among the firms licensed to buy them. After months of export controls, Washington is adjusting the GPU valve. Not fully open, not fully shut: classic semiconductor diplomacy.
Twenty-six Meta employees allege the company used AI systems to target workers on medical, parental or family leave during layoffs. Meta denies the claims and says people made the decisions. The case could become a major test for workplace AI, a phrase that already sounds like a compliance migraine.
TYLSemi has raised $43 million to build open-standard chiplets for custom AI chips. Founded by former Qualcomm and AlphaWave executives, the startup wants to loosen proprietary lock-in around high-speed chip links. In AI infrastructure, even the Lego bricks now need a strategy deck.
Anthropic is launching Claude for Teachers, giving verified US K-12 educators one year of free premium access, curriculum connectors aligned with all 50 state standards, teaching skills, and student-data protections. AI is entering the classroom through lesson prep, not the back row with gum.
The White House wants utilities, data center developers and governors to back a voluntary pledge aimed at keeping AI power demand off household bills. An earlier deal with tech giants was mostly symbolic; this version is looking for actual bolts to tighten.
Meta introduced Muse Spark 1.1, a multimodal reasoning model built for agentic tasks, coding, tool use and visual understanding. It is available in the Meta AI app and through the new Meta Model API in public preview. After Llama, Meta is also putting reasoning on the counter.
More than 200 researchers and economists, including 16 Nobel laureates according to Stanford, are urging governments and companies to prepare now for AI’s economic impact. Their message is simple: waiting for certainty would be like packing the parachute after the jump.
OpenAI has launched GPT-Live, a family of voice models that can listen and speak at the same time. Two versions, GPT-Live-1 and GPT-Live-1 mini, are rolling out globally. The goal is smoother voice agents that interrupt less, which is already a small act of civilization.
Elon Musk says SpaceXAI is making Grok 4.5 public as OpenAI prepares GPT-5.6. Axios describes a model focused on coding and agentic work, positioned as faster and cheaper than top rivals. The Musk-OpenAI rivalry still runs on compute, pricing and a reliable supply of theater.
OpenAI plans to publicly launch GPT-5.6 on Thursday after a limited rollout requested by Washington over national security concerns. Sol will be joined by the lower-cost Terra and Luna models. The frontier model now ships with benchmarks and a fairly visible government clipboard.
The White House is negotiating voluntary standards with several AI labs for releasing advanced models. The framework could define capability thresholds, timelines, and access rules in the US and abroad. An announcement may come as soon as next week if the talks reach agreement.
Two Super Micro employees in Taiwan have been detained in an investigation into the alleged export to China of AI servers containing restricted Nvidia chips. Two others were released on bail. The company says it is cooperating and is not itself a target of the probe.
GitHub will remove its model catalog, playground, inference API, and BYOK option on July 30, including for active customers. Two preparatory brownouts are scheduled for July 16 and 23. GitHub is directing developers to Azure AI Foundry and Copilot.
AWS is committing $1 billion to a Forward Deployed Engineering organization that will embed thousands of engineers with customers, plus $1 billion in credits to modernize US intelligence cloud systems. Secret Cloud is also opening to cleared defense contractors. Agentic AI is putting on a very official badge.
OpenAI has discussed giving the US government a 5% stake to share some of the value created by AI, according to the Financial Times. The proposal remains preliminary. Anthropic says it has held no comparable talks with the Trump administration.
Anthropic has launched Claude Science in beta on macOS and Linux for Pro, Max, Team and Enterprise users. The workspace combines literature, code, compute and more than 60 scientific skills, from genomics to chemistry. Figures, analyses and manuscripts remain reproducible, while a reviewer agent checks citations and calculations.
Anthropic has launched Claude Sonnet 5, now the default for Free and Pro plans and available in Claude Code and the API. It improves reasoning, coding and tool use, at times approaching Opus 4.8 performance for less. Intro pricing is $2/M input tokens and $10/M output tokens through August 31.
US restrictions on Fable 5 and Mythos 5 have been lifted. Anthropic is restoring global Fable 5 access on July 1, with a classifier blocking over 99% of the reported bypass. Mythos 5 is returning first to select US organizations. The regulatory saga has finally found a resume button.
Firmus says it has partnered with Nvidia to deploy 170,000 GPUs in Batam between 2027 and 2028. The Australian company plans to sell Nvidia-powered cloud services to smaller AI firms and targets up to $30 billion in revenue over six years. Affordable compute, apparently, now arrives by the container.
The Bank for International Settlements warns that massive AI investment could create overcapacity and a boom-bust cycle. It also flags growing reliance on debt and complex financing. The five largest hyperscalers are expected to spend more than $1 trillion on AI-related capex across 2025 and 2026.
Google could not supply all the Gemini capacity Meta requested, delaying some internal projects, the Financial Times reports. Meta reportedly told staff to conserve tokens. Reuters could not verify the report, and neither company commented. Even the largest AI buyers are discovering power-saving mode.
The US government is allowing Anthropic to redeploy Mythos 5 to a small group of American cyber defenders and critical infrastructure providers. The model had been suspended on June 12, while Fable 5 remains unavailable. After two weeks of negotiations, Claude has recovered only part of its access badge.
Anthropic’s new Economic Index maps Claude usage rhythms and links them to a survey of about 9,700 users. Personal conversations rise from roughly 35% on weekdays to nearly 50% on weekends. Heavy delegators are also more optimistic about work, though the sample overrepresents tech and management roles.
Mirendil, founded by former Google and Anthropic researchers, has raised a $200 million seed round at a reported $1 billion valuation. The startup aims to give scientists tools for training models that can accelerate AI research itself. It is a seed round wearing the financial outfit of a much later stage.
Featured
OpenAI is starting a limited preview of GPT-5.6 with Sol, Terra and Luna. Sol advances in coding, biology and cybersecurity, with an “ultra” subagent mode. At Washington’s request, access begins with selected partners before a broader rollout. The most capable model arrives with a bouncer at the door.
An internal OpenAI study says Codex now accounts for over 85% of the average employee’s output tokens and 99.8% of weekly output tokens company-wide. Growth is fastest among non-developers. The figures come from OpenAI and partly rely on model-estimated task duration, but they show how far agent-based work has already spread inside the company.
Microsoft released the third edition of its AI in Education Report and announced new tools at no additional cost ahead of ISTELive 26. The company aims to help teachers and students move from experimentation to more structured use, backed by additional training and guidance. Schools are getting homework from AI too.
More than 70% of euro-area firms reported using AI in late 2025, but only 7% used it intensively, according to the ECB. Nearly half of non-users planned to invest in 2026. AI has clearly entered the office, but it is not yet working there full-time.
Nvidia has announced BioNeMo Agent Toolkit, a set of tools for agents specialized in biology and drug discovery. It combines domain-specific tools and skills to orchestrate scientific tasks. The release extends Nvidia’s role beyond compute and deeper into research workflows.
A jobs.ch study says the share of entry-level listings in Switzerland was 32% lower in 2025 than the 2019-2022 average, as companies adopted more AI. Correlation does not prove AI caused the entire decline, but the first rung of the career ladder is clearly becoming harder to find.
Since April, several users had reported massive SSD writes caused by Codex’s local SQLite logs, sometimes reaching several megabytes per second. OpenAI shipped two fixes in Codex 0.142.0 on June 22, cutting log volume by around 85% in the main reporter’s tests.
Featured
Sakana AI has unveiled Fugu, a system that coordinates multiple AI agents behind a single API. Instead of relying on one massive model, Fugu distributes tasks across specialized agents and claims performance comparable to Fable and Mythos. The launch highlights a growing trend in AI: orchestration and coordination may become just as important as the models themselves.
Featured
Midjourney is launching a medical division around “Ultrasonic CT,” a full-body ultrasound scanner. The company says it aims to deliver imaging superior to MRI in several ways, with scans taking as little as 60 seconds and requiring neither radiation nor powerful magnetic fields. Its first center is planned for San Francisco in late 2027.
SpaceX has announced the acquisition of Cursor in an all-stock transaction. The company says its teams have already spent several months jointly training a new AI model that will soon be integrated into both Cursor and Grok Build. The deal further expands SpaceX’s ambitions in AI-powered software development and could position Cursor as a key part of the broader Grok ecosystem.
Featured
Anthropic says it must suspend access to Claude Fable 5 and Mythos 5 following a U.S. government export control directive. The restriction applies to all users, including some foreign Anthropic employees. The company calls it a misunderstanding and says it is working to restore access as quickly as possible.
Moonshot AI has unveiled Kimi-K2.7-Code, a new version of its coding-focused model. Beyond the benchmark gains, one detail stands out: the company is emphasizing agent workflows and the MCP ecosystem, now central to tools like Cursor, Codex, and Windsurf. The model also claims 30% lower reasoning token usage, with a 6x faster mode already on the way.
OpenAI is rolling out a long-requested Codex feature: rate limit resets can now be saved and used later instead of being applied immediately. Go, Plus, Pro, and Business subscribers are getting one free reset to start. A small update, but a meaningful boost in flexibility for users saving their resources for bigger projects.
According to The Wall Street Journal, OpenAI is reportedly considering major price cuts to compete more aggressively with Anthropic, particularly on token costs. Nothing has been confirmed yet, and Reuters says it could not independently verify the report. But the signal is clear: the AI race is no longer only about models, it is also about pricing.
YouTube is testing a new AI-powered tool that can generate thumbnails directly inside YouTube Studio. Creators can describe their idea and receive multiple thumbnail concepts, along with explanations of why they might perform well. After titles, dubbing, and summaries, YouTube continues to automate more of the creator workflow.
A new text-to-text model called celestial-atropos has been spotted on LMArena. These appearances often precede official announcements from major AI labs, sometimes by days, sometimes by several weeks. Its origin remains unknown, but this kind of public testing is usually a strong signal that a major model is on the way. Some are already speculating about OpenAI.
The Information has once again fueled speculation around OpenAI’s next frontier model, which could reportedly launch as early as this month under the GPT-5.6 name. Nothing has been officially confirmed, but a growing number of clues spotted in recent weeks have kept the rumor mill spinning. That said, it’s probably not the boldest prediction of the year: saying OpenAI will release a new model in the coming weeks is a bit like predicting the sun will rise tomorrow morning. At this point, The Information clearly isn’t afraid to take risks.
Featured
Anthropic has released Claude Fable 5, its most capable public model yet. Behind the announcement lies a simple idea: Fable appears to be a lighter and more restricted version of Mythos 5, Anthropic’s frontier model previously limited to a small group of trusted partners. While both models reportedly share the same foundation, Fable includes stronger safeguards around sensitive areas such as cybersecurity and biology. For the public, it may be the closest look yet at the capabilities of a Mythos-class system without full access to the original model.
Featured
After two years of delays, Apple unveiled “Siri AI” at WWDC 2026. The new assistant can understand on-screen content, retrieve information from messages, and maintain conversational context. Apple is positioning the upgrade as its answer to ChatGPT, Gemini, and Claude.
OpenAI is rolling out several ChatGPT updates across web, iOS and Android, including interactive charts, tables of contents for longer conversations, a full-screen editor for long-form writing, and email sending directly from chat. It is not a major AI breakthrough, but it shows where the product is heading: ChatGPT is becoming less of a simple chatbot and more of a complete work interface.
Moonshot AI has launched @KimiDevs, an official X account dedicated to developers building with Kimi Code and the Kimi API. The verified account is expected to become the main channel for product announcements, technical updates, and ecosystem news. Its first post is a simple but symbolic “Hello world!”.
Featured
According to the latest Cloudflare Radar data, automated agents now account for 57.3% of HTML requests, compared to 42.7% generated by humans. It marks the first time in internet history that machine-driven traffic has overtaken human activity online. The shift highlights the rapid rise of AI agents, crawlers, and automated systems across the web.
Miso has unveiled Miso One, an open-source 8-billion-parameter text-to-speech model. According to the company, it delivers significantly more expressive and human-like speech than existing solutions while maintaining just 110ms of latency. The model weights are already available, with API access coming soon. The release could help accelerate the adoption of AI-powered voice agents.
Bloomberg, cited by Reuters, reports that Google is months behind schedule on Gemini 3.5 Pro, its flagship model, as it works to improve capabilities, especially coding. A launch expected in June has reportedly slipped. Even Google can meet the hourglass.
China is considering allowing Alibaba, ByteDance and DeepSeek to buy a limited number of Nvidia H200 chips, according to The Information via Reuters. The total could be below 200,000, less than half earlier requests. AI sovereignty still has to check the GPU inventory.
Google has released Gemini Omni Flash in public preview. It creates 3 to 10-second 720p videos from text or a still image, then edits them through conversation. Google also made Nano Banana Lite generally available for low-latency image generation and editing. The creative studio now fits inside an API.
Featured
Austria is asking Brussels to explore a strategic Anthropic presence in the EU after US restrictions on access to advanced models. Vienna is offering legal certainty, market access and capital, without explaining how the plan would work. European AI sovereignty has moved from speeches to a rather direct invitation.
Kunlunxin, Baidu’s AI chip unit, is preparing a Hong Kong listing at a target valuation of $50 billion, The Information reports. Reuters could not verify the figure and Baidu did not comment. The company had already confirmed a confidential listing application in January.
OpenAI is making Codex Remote generally available across all ChatGPT plans. From the mobile app, users can start or resume work on a connected Mac or PC, monitor progress and approve actions. Pairing now uses an authenticated one-to-one QR code, keeping the agent’s remote work under at least some supervision.
OpenAI, Anthropic, Microsoft and Amazon are backing RAISE US, an organization designed to prepare American workers for AI-driven disruption. More than $500 million has been committed, with pilots in four states covering training, career guidance and new support mechanisms. The industry is funding the post-automation plan before finishing the automation.
General Intuition has raised a $320 million Series A to build models that can perceive, predict and act across virtual and physical environments. TechCrunch reports a $2.3 billion valuation, with gaming data playing a central role. After learning to talk, AI models now want a decent sense of space.
Axios and the Financial Times report that the Trump administration asked OpenAI to limit GPT-5.6’s first release to approved partners while security risks are assessed. OpenAI and the White House have not commented. The next model launch may look less like a queue and more like a guest list.
Steve Jarrett, Orange’s chief AI officer, is leaving the telecom group to join Anthropic. Based in Paris, he will support the Claude maker’s expansion across Europe and Africa. The hire gives Anthropic deep enterprise and telecom experience. Claude has found a new local guide, complete with a well-stocked address book.
The European Commission is joining Pax Silica, the US-led initiative focused on AI supply chains. The pact spans energy, critical minerals, semiconductors, advanced manufacturing and AI models. Brussels is betting on greater resilience, although Europe’s technology sovereignty now comes with a noticeable American accent.
Representative Nathaniel Moran has introduced the AI Incident Reporting Act. Frontier model developers would have seven days to notify the Commerce Department of security breaches, loss-of-control behavior or dangerous capabilities. The most serious cases would reach Congress within 48 hours. AI safety may finally get a formal alarm button.
Featured
OpenAI and Broadcom unveiled Jalapeño, an accelerator built specifically for large-model inference. Engineering samples are already running GPT-5.3-Codex-Spark workloads. OpenAI targets initial deployment by late 2026 and says the chip should cut cost, latency and power use, though final performance figures are still pending.
Anthropic says operators linked to Alibaba and Qwen used nearly 25,000 fake accounts for 28.8 million Claude interactions between April and June. The alleged goal was to distill its agentic reasoning and coding capabilities. Alibaba has not publicly responded to the claims.
Meta is launching a new range of AI-powered smart glasses starting at $299. The company is expanding its push beyond models and apps into a consumer device designed to be worn throughout the day. The assistant race may end up being fought on our faces, a place where notifications had so far shown admirable restraint.
Anthropic has launched Claude Tag in beta for Team and Enterprise customers. Inside Slack, teams can mention @Claude, grant access to selected channels, tools and codebases, and delegate tasks in a shared thread. The agent keeps relevant context and can plan work to complete later.
Featured
Mistral has introduced OCR 4, a document-understanding model supporting 170 languages. It returns text, bounding boxes, block types and confidence scores, with a self-hosted option. Mistral positions it for RAG pipelines, enterprise search and automated document processing.
Agility Robotics, maker of the Digit humanoid, plans to merge with SPAC Churchill Capital XI, according to the Wall Street Journal and Reuters. The deal would value the company at about $2.5 billion and could raise more than $600 million. Humanoids are now learning to walk toward Nasdaq too.
The Trump administration is pressing Meta to accept voluntary government reviews of its frontier AI models before release, according to the New York Times via Reuters. OpenAI, Anthropic, Google DeepMind, Microsoft and xAI already take part in related evaluations. Meta says it shares the goal and hopes to sign soon. Voluntary, with a government pen resting nearby.
The UN secretary-general is calling on AI companies to disclose more data on energy and water use, as well as emissions. As data centers multiply, the true footprint of AI models remains difficult to compare. AI is excellent at measuring the world, but still rather shy when asked to measure its own environmental bill.
Nvidia says its technology powers 81% of the world’s 500 fastest supercomputers and 90% of the systems newly added to the list. The top eight Green500 machines also use Nvidia GPUs. The figures underline the company’s reach across scientific computing and AI infrastructure, far beyond generative models alone.
OpenAI has published a case study in which GPT-5 Pro helped immunologist Derya Unutmaz and his lab revisit a long-standing puzzle involving immune cells linked to cancer and other diseases. The example shows how models can support scientific analysis while leaving validation to researchers.
GitHub has joined Black Forest Labs, Hugging Face and Mozilla in calling for targeted changes to California’s AI Transparency Act. The coalition argues that proposed license-revocation rules conflict with the perpetual and irrevocable nature of open-source licenses. A technical dispute with potentially broad consequences.
The illustrations and animations used for Claude Fable 5 were not generated with AI. According to Sam McAllister, they were created from Brockhaus encyclopedia archives using Photoshop, After Effects, and external animators. An unusual choice for one of the world's leading AI companies. Curious, for a team that promises to automate creative work.
Anthropic says the world should have the option to slow down, or temporarily pause, the development of the most advanced AI models if risks become too high. The core concern is AI systems becoming capable of contributing to their own improvement. But a credible pause would require major labs, including in the US and China, to agree on verifiable rules. In other words, Anthropic is raising a real governance issue, but also a very hard enforcement problem.
OpenAI has confidentially submitted an S-1 filing to the SEC, marking a first step toward a potential IPO. The company says no timeline has been decided yet and that staying private may still make sense for now. The message is clear: the AI race is no longer only about models, but also about access to capital.
A vulnerability in Claude Code GitHub Action allowed attackers to exploit a simple GitHub issue through prompt injection to exfiltrate CI/CD secrets and gain elevated permissions on certain repositories. Reported by RyotaK, the flaw has been patched by Anthropic, but it highlights the risks of exposing AI agents to untrusted content.
