The top things worth knowing about in AI today.
OpenAI put GPT-5.6 into public release on 9 July across ChatGPT, Codex and its API, split into three price tiers: Sol ($5/$30 per million tokens), Terra ($2.50/$15) and Luna ($1/$6). It also launched ChatGPT Work, an agent meant to complete whole tasks across a team's documents and tools rather than just answer questions. OpenAI says Sol beats Anthropic's Fable 5 on coding while using less than half the output tokens, so the practical story is cost and speed as much as capability.
Read moreMicrosoft launched Frontier Company, a $2.5 billion business that embeds 6,000 engineers and industry specialists directly inside client organisations to build and run their AI systems. Early customers include the London Stock Exchange Group, Unilever, Novo Nordisk and Land O'Lakes. The bet is that the hard part of enterprise AI is now deployment, not the models, a useful signal for any business struggling to turn pilots into results.
Read moreAnthropic agreed a 20-year lease with TeraWulf for a Kentucky data centre campus expected to generate around $19 billion in revenue, built on a former aluminium smelter site. The campus reaches 401MW of capacity by early 2028, with first power in late 2027. The deal alone exceeds TeraWulf's market value, showing how much compute the leading labs are locking in years ahead of demand.
Read moreCloudflare now splits web crawlers into search, agent and training categories and lets site owners control each separately, blocking training and agent bots on ad-supported pages by default from 15 September. Cloudflare sits in front of roughly a fifth of the web, and the controls are on every plan including free. If you run a website, you now decide whether AI firms can take your content to train models or answer questions in chatbots.
Read moreThe US Federal Trade Commission proposed a policy statement saying AI companies that quietly steer model outputs toward undisclosed goals could be deceiving consumers under Section 5 of the FTC Act. Firms can prioritise objectives other than accuracy, but only if they disclose it clearly, and complying with a state law is not a defence. Comments run until 31 July, putting model behaviour and transparency on the regulatory agenda.
Read moreMistral released Leanstral 1.5, a free Apache-licensed model built for Lean 4, the tool used to mathematically prove that software behaves as intended. Tested across 57 code repositories it flagged 47 broken properties and 11 genuine bugs, five of them never reported before. For teams working on critical systems, that points to AI that can formally verify code rather than only write it.
Read moreGoogle added Video Remix to Google Photos, using its Gemini Omni model to restyle clips with relighting, background swaps and effects such as watercolour or oil painting from a template. It is live for paying Google AI subscribers in the US, India, Japan and several other markets, capped at 10-second clips that take up to two minutes to generate. It shows where consumer video editing is heading: describe the change, skip the software.
Read moreAssistant science and technology minister Andrew Charlton told an AI safety forum in Sydney that powerful models are already doing things their creators never intended, as Australia stands up an AI Safety Institute to test models and advise regulators. Its budget of A$29.4 million over four years is dwarfed by the UK equivalent of roughly A$460 million. For Australian businesses it means local oversight is coming, but light-touch for now.
Read moreA short read every week — the few things worth your time, and nothing that isn’t.