Models, agents, safety, research and the companies building them. We cover verifiable developments in AI with primary sources and clear attribution.
Featured Story
Anthropic introduced Claude Sonnet 5, positioning the model for coding, tool use and professional work while publishing safety-evaluation caveats and launch pricing.
Latest · 78 stories
HP and OpenAI announced a strategic partnership to move Frontier from pilots into customer support, telemetry, employee productivity and software-development workflows.
OpenAI published an EU-focused jobs transition framework using ESCO and Eurostat data to examine how AI capabilities may affect occupations across member states.
Reports say OpenAI and Anthropic limited access to new frontier systems during a US government cybersecurity review, putting model launches inside a more political approval process.
Google says developers can now use Gemini 3.5 Flash as a native computer-use agent across browser, mobile and desktop environments.
Claude Tag lets Enterprise and Team customers bring Claude into selected Slack channels with scoped tools, channel memory, logs and spend limits.
OpenAI says Daybreak combines Codex Security, GPT-5.5-Cyber and partner workflows to help defenders validate vulnerabilities, generate fixes and move security work closer to deployment.
OpenAI says Ona will bring secure, customer-controlled cloud environments into Codex, sharpening the enterprise question around where long-running agents work and who governs them.
OpenAI says GPT-5.5 Instant improves health-answer quality for free ChatGPT users, citing physician-led evaluations and production monitors, but the evidence remains company-run and not clinical validation.
Google DeepMind says agent deployment needs system-level controls, live monitoring and response plans that assume even aligned models can behave unexpectedly.
A Google DeepMind prototype built with Gemini aims to help UK planning officers process householder applications faster, turning AI into a public-service workflow experiment.
Anthropic says Claude Corps will train 1,000 fellows, place them with more than 400 nonprofits and commit an initial $150 million to widen AI benefits during workplace disruption.
NIST highlighted a peer-reviewed mathematical proof arguing that finite AI guardrail sets cannot be universally robust against adaptive adversarial prompts.
OpenAI introduced usage analytics and updated spend controls for ChatGPT Enterprise, giving administrators a clearer view of credit consumption across users, products and models.
OpenAI says a deployment-simulation method can replay de-identified real conversation contexts through candidate models to estimate some undesired behaviours before release.
Google says its June Pixel Drop is rolling out AI-assisted video creation, conversational photo editing and expanded voice translation, making generative features a device-level consumer workflow rather than a separate app destination.
OpenAI says it will invest $150 million in a new global partner ecosystem, aiming to train 300,000 certified consultants and move enterprise customers from pilots toward production deployments.
The 26B mixture-of-experts model uses diffusion-style parallel generation rather than standard next-token decoding, with Google describing it as experimental and best suited to speed-critical local workflows.
NSPM-12 sets new governance principles for National Security Systems, re-establishing CNSS structures and formalising the NSA Director's role as National Manager for NSS cybersecurity.
Anthropic says it disabled access to two frontier models after a US export-control directive, while disputing that a narrow jailbreak finding should justify recalling a deployed commercial system.
OpenAI says it has terminated accounts connected to covert influence campaigns originating in China, including operations that used its tools to generate multilingual content and translate political messaging.
The European AI Office has published updated drafts of the codes of practice that will govern general-purpose AI models under the EU AI Act, with provisions covering transparency, copyright and systemic-risk evaluation.
Anthropic has revised its published model specification to include guidance on how Claude should behave when operating as an agent, receiving instructions from other AI systems, and acting with greater autonomy in multi-step tasks.
The UK regulator's latest AI Airlock phase argues that medical AI oversight cannot stop at pre-market validation, especially when model behaviour, human oversight and clinical context can change after deployment.
Anthropic says its newest generally available model can handle longer, more complex work, while a less restricted version will remain limited to approved cybersecurity organizations.
Google said Gemini 3.5 Live Translate can deliver near real-time speech-to-speech translation across more than 70 languages.
The company has confidentially submitted a draft registration statement, but says it has not decided when, or whether, to proceed with an IPO.
Alphabet published a June investor presentation after announcing a proposed equity raise to support AI infrastructure and growth demand.
OpenAI outlined 2026 election safeguards covering reliable voting information, cyber-defender support, transparency around AI-generated content and misuse monitoring.
Google’s I/O recap listed Gemini 3.5 Flash, AI Studio, Antigravity and agentic developer tools among the headline announcements.
At I/O 2026, Google said it was upgrading Search with Gemini 3.5 Flash, agent-style features and a redesigned AI-powered search box.
OpenAI launched a deployment-focused company and agreed to acquire Tomoro, saying the new unit will embed forward deployed engineers into organisations building production AI systems.
Anthropic released ten ready-to-run agent templates for financial services, including pitchbooks, KYC screening, valuation review and month-end close workflows.
The Chief Digital and Artificial Intelligence Office listed classified-network AI agreements dated 1 May, signalling continued federal movement to place advanced AI tools inside secure national-security environments.
OpenAI proposed policy ideas around access, prosperity and institutional resilience as AI capability advances.
OpenAI launched ChatGPT Images 2.0, presenting the update as a new era for image generation inside ChatGPT.
Anthropic signed a new agreement for multiple gigawatts of next-generation TPU capacity to support future Claude models.
Google’s April AI recap highlighted Gemma 4, Deep Research Max and Cloud Next announcements around enterprise agents.
Google said Gemma 4 was its most capable open model family to date, with agentic workflows, multimodal inputs and commercially permissive licensing.
OpenAI said it had closed committed capital at an $852 billion post-money valuation, underscoring how frontier AI has become an infrastructure-scale financing race.
AWS said its 2026 AI and ML Scholars programme would provide free AI education, with top performers eligible for a funded Udacity nanodegree.
OpenAI published a detailed account of how it writes and updates the public framework that guides model behaviour.
Google said Personal Intelligence was expanding in the US, connecting user context from apps such as Gmail and Photos to tailor AI responses.
OpenAI said it will acquire Promptfoo and integrate its AI security-testing technology into Frontier for enterprise agent deployments.
OpenAI says GPT-5.4 is its most capable and efficient frontier model for professional work, with GPT-5.4 Thinking in ChatGPT and availability in the API and Codex.
OpenAI and Amazon announced a multi-year partnership to bring OpenAI Frontier to AWS customers and use Trainium compute for growing demand.
Anthropic says Sonnet 4.6 improves coding, long-context reasoning, agent planning and resistance to prompt injection while adding a 1M-token context window in beta.
OpenAI updated deep research so users can connect apps through MCP, restrict web searches to trusted sources and track progress in real time.
A new index surveyed public servants across 10 countries, measuring not only whether governments use AI but whether staff feel enabled to use it responsibly.
The OECD Digital Education Outlook 2026 put generative AI in education under a practical lens: better output does not automatically mean deeper learning.
Anthropic said ServiceNow selected Claude to power customer applications and increase internal productivity.
Anthropic released a new constitution describing desired Claude behaviour, values and operating context.
NIST’s CAISI issued a request for information on securing AI agent systems as agentic tools move into real workflows.
OpenAI announced a dedicated health and wellness experience in ChatGPT, built with physicians and framed around privacy and safety.
NIST published research using AI-driven antimicrobial peptide characterisation to identify motifs relevant to drug design.
OpenAI marked more than one million business customers, framing 2025 as the year enterprise AI moved from experimentation to operational use.
NIST released a preliminary draft Cybersecurity Framework Profile for AI to help organisations secure, defend and thwart AI-related cyber risks.
OpenAI donated AGENTS.md and joined the Linux Foundation effort to create neutral standards for agentic AI.
Anthropic acquired Bun and said Claude Code had reached a $1 billion annualised run-rate milestone.
Anthropic launched Claude Opus 4.5, describing it as its strongest model for coding, agents and computer use.
OpenAI described GPT-5.1-Codex-Max as a coding model built for sustained agentic software tasks.
OpenAI introduced GPT-5.1, describing the update as more conversational, more customizable and more efficient for developers.
OpenAI introduced IndQA, a benchmark designed to evaluate AI systems on Indian culture and languages.
OpenAI announced Aardvark, a GPT-5-powered agent designed to find, validate and help fix software vulnerabilities.
OpenAI introduced ChatGPT Atlas, a macOS browser with ChatGPT built into the browsing interface and agent features for paid users.
Anthropic launched Claude Haiku 4.5, positioning the small model around near-frontier coding quality at lower cost and latency.
OpenAI introduced apps in ChatGPT and a preview Apps SDK at DevDay, turning ChatGPT into a distribution surface for third-party services.
Anthropic introduced Claude Sonnet 4.5, positioning the model around coding, computer use, reasoning and longer agentic work.
NIST presented work on adapting SP 800-53 security controls for AI-specific risks such as model integrity, data provenance and adversarial robustness.
The OECD's Governing with Artificial Intelligence report examined how governments can use AI while managing bias, transparency, capability and public-trust risks.
Anthropic launched a preview that lets Claude create and edit spreadsheets, documents, slide decks and PDFs directly in Claude.ai and the desktop app.
UNESCO's report on AI and education argued that generative AI in schools must be governed around equity, privacy, safety and human-centred learning.
The White House released America's AI Action Plan, framing AI leadership around innovation, infrastructure build-out and international technology strategy.
Commission guidelines explained how AI Act obligations apply to providers of general-purpose AI models, including open-source exemptions and market-placement rules.
The European Commission published the General-Purpose AI Code of Practice, a voluntary tool for providers preparing for AI Act obligations.
Texas Governor Greg Abbott signed the Texas Responsible Artificial Intelligence Governance Act on 22 June 2025, adding another major state-level AI compliance regime in the United States.
Anthropic's 22 May 2025 Claude 4 launch focused on coding, reasoning and agent workflows, with Claude Code reaching general availability alongside new API capabilities.
At Google I/O 2025, Google said AI Mode would move into a dedicated Search tab, turning complex, multi-step queries into a mainstream product surface rather than a lab feature.
Google DeepMind introduced Veo 3, Imagen 4 and Flow on 20 May 2025, making video, image and filmmaking tools a more explicit part of the AI platform race.
AI Coverage
Subscribe on the homepage to be notified when the newsletter launches — evidence before excitement, delivered to your inbox.
