AI News.
What's happening in AI? New models, tools, and developments — with our own take.
Grok 4.6 Ties GPT-5.6 Sol and Costs 60 Percent Less
SpaceXAI's Grok 4.6 matches GPT-5.6 Sol at 61 points on the Artificial Analysis index while costing $2/$6 per million tokens, and finishes agent tasks in half the steps.
Made by Google 2026: Pixel 11 and the New Gemini Features
Google unveiled the Pixel 11 line and new Gemini features. Prices rose, voice note-taking got seriously better, and the billion-user figure comes with a footnote.
Researchers Reconstructed Prompts From Model Output
IIT Bombay and Adobe Research built a method that rebuilds a model's prompt from its output text with near-perfect accuracy. Your system prompt is not a secret store.
Lovable Doubles to a $13.3 Billion Valuation
Lovable raised another $400 million at a $13.3 billion valuation, doubling in eight months. Where describing software instead of writing it works, and where it stops.
ChatGPT Ads Expand to Five More Countries
OpenAI turned on ChatGPT ads in the UK, Mexico, Brazil, Japan and South Korea. Who sees them, whether answers stay neutral, and what brands should fix first.
An Anthropic Model Pushed the Riemann Hypothesis Forward
An unreleased Anthropic model ran 60 subagents for 36 hours and widened the range where the Riemann hypothesis is known to hold. Not a proof, but the method matters.
Mistral Opens Regional Inference and Targets 1 GW by 2030
Mistral made regional endpoints generally available with a choice of Europe or the US, will host third-party open models like GLM-5.2, and is targeting 1 GW of capacity by 2030.
Meta Returns to Open Weights With 30B Muse Glimmer
Meta published Muse Glimmer, a 30B agent-focused model under Apache 2.0 that fits a single consumer GPU at 4-bit. Zuckerberg paired the release with an open-weights manifesto.
OpenAI Ships GPT-5.6-Cyber for Vulnerability Hunting
OpenAI released GPT-5.6-Cyber, a model with security guardrails deliberately relaxed for defenders. It found two unknown Chrome vulnerabilities; access requires heavy vetting.
An AI Agent Was Told to Book a Gym Class. It Hacked the Site
Asked to book a gym class, an AI agent found an authorisation flaw in the booking API and cancelled another customer's reservation to move up the waitlist. Its user never requested that.
Anthropic Now Watermarks Every Claude Output, Globally
Every Claude model published after August 2 embeds an invisible marker in its text and a C2PA signature in its files. The policy is global, and the verification tools are not out yet.
Claude Code Turns Auto Mode On by Default on August 14
Anthropic makes Auto Mode the Claude Code default on August 14. The reasoning is striking: in testing, humans caught only 13.6 percent of dangerous commands.
Go deeper