The top things worth knowing about in AI today.
OpenAI has started a limited preview of GPT-5.6, a three-tier family of Sol, Terra and Luna, available first through its API and Codex to trusted partners before a wider release. Sol, the flagship, is priced at $5 per million input tokens and $30 per million output, with stronger results in software engineering, scientific research and cybersecurity. A broad ChatGPT rollout was expected this month but has slipped, so most users will not see it until July.
Read moreOpenAI and Broadcom unveiled Jalapeno, OpenAI first custom chip built specifically for running AI models rather than training them. OpenAI says it designed the accelerator in about nine months, using its own models to speed up the work, and frames it as the first in a multi-generation hardware platform built with Broadcom. Designing its own silicon is OpenAI bid to cut its dependence on Nvidia and lower the cost of serving models at scale.
Read moreThe Commerce Department has allowed Anthropic to release its most capable model, Claude Mythos 5, to about 100 named US companies and federal agencies. The government had suspended access to Mythos 5 and Fable 5 on 12 June, citing national security, blocking use by any foreign national including Anthropic own staff. The partial relief, set out in a letter from Commerce Secretary Howard Lutnick, lets approved organisations use the model without an export licence.
Read moreThe European Parliament has approved amendments that push back the EU AI Act toughest obligations. Rules for standalone high-risk systems in areas such as hiring, education and credit scoring now apply from December 2027 rather than August 2026, a 16-month delay. For Australian firms selling into Europe it buys time, but the underlying duties, including risk assessments and documentation, still arrive.
Read moreAlphabet is raising about $80 billion in equity to fund its AI build-out, including a roughly $10 billion investment from Warren Buffett Berkshire Hathaway. The raise supports capital spending heading toward $190 billion this year, most of it on data centres and chips. The scale shows how far the largest players will go to secure compute, even as investors question how quickly the spending will pay off.
Read moreNew industry data shows agentic AI moving into production across finance, healthcare and retail, with 72% of surveyed firms running agents live, but 60% lack formal governance for them. Only 34% of organisations say their AI programs produce measurable financial impact, and fewer than 10% have scaled agents beyond pilots. The gap between deployment and disciplined oversight is now the main barrier, not the technology itself.
Read moreOn OpenRouter, which routes developer traffic across many AI providers, DeepSeek models now process more tokens than Google and OpenAI combined, at about 17.6% of weekly volume. Chinese-built models together account for roughly 46% of identified usage on the platform. The shift reflects how many cost-conscious developers are choosing cheaper open models for production work.
Read moreA short read every week — the few things worth your time, and nothing that isn’t.