The top things worth knowing about in AI today.
Google pushed its rebuilt Gemini 3.5 Pro to 17 July after scrapping the previous architecture over weak coding and reasoning results. The delay followed four senior DeepMind researchers leaving in a single week — including Gemini co-lead Noam Shazeer to OpenAI and Nobel laureate John Jumper to Anthropic — and roughly $225 billion wiped off Alphabet's market value. The launch will test whether Google can still match GPT-5.6 and Claude at the frontier.
Read moreA CNBC investigation found Chinese open-weight models such as DeepSeek and Z.ai's GLM now account for up to 46% of enterprise token usage on the OpenRouter platform, up from about 11% a year ago. The draw is price: these models run 60 to 90% cheaper than OpenAI or Anthropic, with DeepSeek V4 Flash at $0.14 per million input tokens against $5 for GPT-5.5. For teams weighing AI costs, the cheaper option is now often a Chinese model.
Read moreMeta launched Muse Spark 1.1 on 9 July through its new Meta Model API — the company's first public API and the first model it charges to use. The multimodal, agentic model carries a 1-million-token context window and is priced at $1.25 and $4.25 per million input and output tokens, roughly a quarter of Claude Opus and GPT-5.6 rates. The preview is US-only for now, with $20 in free credits per new account.
Read moreAnthropic has passed OpenAI on annualised revenue, reaching a run rate above $30 billion against OpenAI's roughly $24 to 25 billion, driven by enterprise and API demand for Claude. The company now carries a $965 billion valuation, above OpenAI's $852 billion. The gap reflects two strategies: Anthropic sells mostly to businesses and developers, while OpenAI leans on consumer ChatGPT subscriptions.
Read morePrime Minister Anthony Albanese today announced an Office of AI inside the Department of the Prime Minister and Cabinet to run a single national framework covering AI's effects on energy, copyright, jobs, education and national security. The government says a unified approach will give investors clearer, faster approvals and one process for checking compliance. Canberra is positioning it as a world-first attempt to pull fragmented AI policy into one place.
Read moreFrom today, China's Interim Measures for anthropomorphic AI services force ByteDance and Alibaba to switch off personalised agents on Doubao and Qwen rather than rebuild them to comply. The rules require anti-addiction systems, usage warnings and instant-exit options that sit awkwardly with agents designed to remember users across sessions. Qwen users lose their agent settings and chat histories permanently, while Doubao keeps read-only access open until 15 October.
Read moreSecurity firm Sysdig published analysis of JADEPUFFER, which it assesses as the first ransomware attack run end to end by an AI agent with no person at the keyboard. The agent broke into an exposed Langflow server, harvested credentials, moved through the network and encrypted 1,342 configuration records — recovering from a failed login to a working fix in 31 seconds. Sysdig's takeaway is blunt: the skill needed to run ransomware has dropped to the cost of running an agent.
Read moreA short read every week — the few things worth your time, and nothing that isn’t.