The top things worth knowing about in AI today.
Google has shown Gemini 4 Argon, its first new frontier model since Gemini 3. It can produce up to 1 million tokens of output, up from 64,000, and scores 77.9% on the DeepSWE software engineering benchmark. Introductory pricing is US$2 per million input tokens and US$10 per million output. Cyber defenders get access first, so most developers and businesses will have to wait for a wider release.
Read moreAnthropic has launched the Claude Frontier Academy, a US$100 million program to train 10,000 engineers to deploy Claude inside large organisations by the end of 2027. It is modelled on a medical residency: in-person training, then a 12-week rotation leading a real project. Commonwealth Bank of Australia is in the first cohort alongside Accenture, Deloitte and McKinsey. The bottleneck in enterprise AI is now people who can ship it, not access to models.
Read moreAustralia's first royal commission into AI began on 1 October in South Australia, led by former Federal Court judge Iain Ross. Machine learning experts Kate Pounder and Professor Christopher Manning sit as commissioners. The inquiry covers jobs, education, public services and regulation, and its final report is due by July 2027. Organisations in SA should expect calls for submissions.
Read moreApple will require more explicit user action before any macOS app gets Full Disk Access, which lets software read files, mail, messages and browsing history. Apple pointed to growing risk from autonomous AI agents, following reports that Meta's Muse app read private messages and a flaw found in ChatGPT's Mac app. Expect desktop AI agents to ask for permissions more often, and check which apps already have this access on your Mac.
Read moreMicrosoft's 2026 Digital Defense Report finds attackers using AI for reconnaissance, social engineering, exploit development and post-breach activity. Government agencies were the most targeted sector, at 27% of observed activity, up from 17% last year. Microsoft also says AI code analysis helps defenders find vulnerabilities earlier. Patch speed and phishing awareness matter more now that attackers move faster.
Read moreOpenAI and Synopsys have announced GPT-Synopsys, a model trained to operate Synopsys chip design tools. Engineers can set goals like power and performance targets, and agents run the tools and iterate on the results. Customer design data is encrypted and excluded from training. It is an early sign that frontier labs are moving into specialised industry models built with incumbent software vendors.
Read moreAnthropic, Google and OpenAI are setting up an industry-run body to set AI safety standards, aiming to launch by early 2027. It would support third-party testing before release, define incident reporting and set qualifications for model auditors. Critics warn it could shut out open-source and smaller developers. For buyers, it may become a shared benchmark for checking vendor safety claims.
Read moreFieldAI is reportedly raising US$700 million at a US$10 billion valuation, five times its 2025 price. Its models let robots navigate without pre-built maps, GPS or internet, and they run on machines such as Boston Dynamics' Spot. Contracts now top US$135 million across construction, energy and public sector customers. Investors are moving money from chatbots towards physical AI in industrial settings.
Read moreA short read every week — the few things worth your time, and nothing that isn’t.