The top things worth knowing about in AI today.
Anthropic released Claude Sonnet 5 on 30 June and made it the default for Free and Pro users. It performs close to the flagship Opus 4.8 on many agentic tasks while costing far less, at $2 per million input tokens and $10 per million output through 31 August, below the old Sonnet 4.6 rate. For most everyday work, the cheaper tier is now good enough that reaching for the top model is harder to justify.
Read moreThe Digital Transformation Agency signed a five-year whole-of-government agreement, VSA6, giving federal agencies access to Microsoft 365, Azure, Copilot and Dynamics 365 from 1 July. It includes capped price increases and an A$1.55 million training fund covering ethical AI. A 2024 Copilot trial found participants saved about an hour a day, the benchmark the rollout is betting on at scale.
Read moreThe US Commerce Department lifted export controls on Anthropic's Fable 5 and Mythos 5 models, which Washington had suspended a month earlier over security concerns. Fable 5 came back for users globally on 1 July, with Anthropic agreeing to detect security risks proactively and report malicious activity. The episode shows how frontier labs now negotiate directly with governments over who can use their most capable models.
Read moreThe intelligence agencies of Australia, the US, UK, Canada and New Zealand issued a joint advisory saying frontier AI models could outpace current cyber defences within months, not years. AI lowers barriers for malicious actors and increases the speed and complexity of attacks, they wrote, urging organisations to patch old systems and tighten access now. The same tools can help defenders spot vulnerabilities earlier, if security teams adopt them first.
Read moreCalifornia is moving its Poppy assistant to a statewide rollout in July after a pilot with more than 2,800 employees across 67 departments, making it the largest US public-sector AI deployment. Poppy runs on the state's own servers, draws only on official CA.gov data, and keeps queries inside government infrastructure. It is a template for how large organisations can adopt AI without sending sensitive data to public services.
Read moreMeta is exploring a cloud business, internally called Meta Compute, to rent out AI models and raw compute to other companies, Bloomberg reported. The move follows roughly $145 billion in AI infrastructure spending this year and continued struggles to get its own models to compete. It mirrors xAI, which began renting out its data-centre capacity to Anthropic after Grok fell behind.
Read moreGoogle released Gemini 3.1 Flash-Lite Image, nicknamed Nano Banana 2 Lite, on 30 June, generating images in about four seconds at a flat $0.034 per 1,000 images. It is available immediately through Google AI Studio, the Gemini API and the enterprise platform. The pricing puts high-volume image work such as product shots, marketing variants and mock-ups within reach of routine business workflows.
Read moreBaseten, which runs the infrastructure that serves AI models in production, raised a $1.5 billion Series F, its fourth round in 18 months. The money is flowing to the unglamorous layer that keeps AI apps fast and reliable rather than to model-building itself. It is a sign investors expect the next wave of returns from deployment and serving, not just frontier research.
Read moreA short read every week — the few things worth your time, and nothing that isn’t.