What's New in AI: Bigger Models, Tougher Rules — 20 Jul

Two stories ran side by side in AI this week, and they pulled in opposite directions. The models kept getting bigger and cheaper — Google shipped its rebuilt Gemini 3.5 Pro and China's Moonshot AI released the largest open-weight model ever built — while governments and institutions moved to set the terms of use. Australia stood up its own Office of AI, France warned that three companies already control the emerging market for AI agents, and Anthropic put Claude directly into classrooms. Here's your plain-English roundup of what happened and why it matters for your business.

Google ships its rebuilt Gemini 3.5 Pro

After a delay and what Google DeepMind described as a ground-up rebuild, Gemini 3.5 Pro arrived around 17 July, timed to coincide with the World AI Conference in Shanghai. The headline features are a two-million-token context window — roughly double what most rivals offer — and a "Deep Think" reasoning mode aimed at complex, multi-step problems and long-horizon coding. Google reportedly scrapped the earlier architecture after engineers found structural weaknesses in areas like recursive tool-calling, then rebuilt from scratch.

Why it matters: This is Google's answer to the public releases of OpenAI's GPT-5.6 and xAI's Grok 4.5, continuing the price-and-capability war we covered in last week's frontier-model roundup. A genuinely usable two-million-token window changes what's practical: you can hand a model an entire contract set, codebase or research archive in one go rather than stitching it together. One caveat worth noting — at launch several specifics circulated from reporting ahead of Google's full model card and pricing, so treat exact benchmark and cost claims as provisional until the official documentation lands. Read more via Tech Times.

Moonshot's Kimi K3 is the largest open-weight model ever

On 16 July, Chinese lab Moonshot AI released Kimi K3, a 2.8-trillion-parameter mixture-of-experts model it calls the biggest open-weight system ever built. It went live immediately through Kimi's apps and API, with the full weights scheduled to publish by 27 July under a modified-MIT licence. Kimi K3 took the top spot on the Frontend Code Arena benchmark with 1,679 points — edging past Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol on that test — though Moonshot's own figures show it still trails those two on overall performance. Launch pricing sits at roughly US$3 per million input tokens and US$15 per million output tokens.

Why it matters: Open weights at this scale mean any company with the hardware can, in principle, self-host a near-frontier model rather than renting one from a US provider. That's a meaningful shift for organisations with data-sovereignty or cost concerns — though a 2.8-trillion-parameter model still demands serious infrastructure to run. It also underlines how fast Chinese labs are closing the gap despite chip export controls. Details via CNBC.

Australia stands up an Office of AI

On 15 July, Prime Minister Anthony Albanese announced a national AI framework and, effective the same day, an Office of AI inside the Department of the Prime Minister and Cabinet to coordinate policy across government. The centrepiece is a set of Australian Standards for AI focused on large data centres: operators would face a legal obligation to underwrite their own new power supply, pay their full share of connection costs so household energy bills aren't affected, reduce load when the grid is stressed, and minimise water use. The government also pledged that Australian writers, artists and journalists will retain ownership of their work, so companies can't train on it without the creator's control. The approach goes to National Cabinet in August, with legislation expected early next year.

Why it matters: This is the local story with the most direct bearing on Australian businesses. The government is signalling a "manage, don't ban" stance — clearer rules, faster approvals, and a social licence for AI infrastructure rather than a moratorium. For organisations planning to build or deploy AI here, the direction of travel is now clearer, and it's the kind of groundwork that makes building and putting AI to work inside a business more predictable. Read the release from the Prime Minister's office.

Anthropic puts Claude in the classroom

On 14 July, Anthropic launched Claude for Teachers, giving verified US K-12 educators a full year of free access to its premium models, Claude Code and Cowork, plus a library of teaching-specific skills. The product connects to Learning Commons for standards mapped across all 50 states, and to curricula like OpenSciEd and Illustrative Mathematics, so lesson plans come out scaffolded and standards-aligned. Anthropic says teacher data won't be used to train its models, and it's working with the American Federation of Teachers on privacy standards designed to meet US student-privacy law.

Why it matters: The race to embed AI in education is now a front in the broader platform contest, and Anthropic is betting on teachers rather than students. The pattern — free access, deep workflow integration, tight data controls — is the same playbook vendors use to win any professional user, and it's worth watching if you evaluate AI tools for your own team. If you're mapping the wider landscape, our curated library of AI tools and resources is a useful starting point. More detail in Anthropic's announcement.

France warns three firms already control AI agents

France's competition regulator, the Autorité de la concurrence, published its opinion on the AI-agents market on 17 July (Opinion 26-A-05), and the standout figure is stark: OpenAI, Google and Anthropic together hold roughly 84% of it. The regulator ran an unusually hands-on inquiry — it built its own AI agents, put 550 shopping questions to them, and logged which sites they visited and cited. Its conclusion is that the shift from chatbots to autonomous agents that can browse, decide and buy on your behalf could concentrate the digital economy around a handful of vertically integrated firms unless regulators act on data access, interoperability and default placement.

Why it matters: Agents are moving from demo to daily use, and whoever owns the default agent owns a powerful gateway to commerce and information. For businesses, the practical takeaway is to avoid locking everything into a single provider's agent stack too early, and to keep an eye on how "agentic commerce" reshapes how customers find you. Read the Autorité's opinion.

OpenAI pushes GPT-5.6 deeper into the office

OpenAI spent the week sharpening ChatGPT for work rather than shipping a new model. According to its release notes, it rolled out a "Work" experience powered by GPT-5.6 to turn scattered notes and drafts into finished output, added desktop updates that make it easier to move between Chat and Work across web, mobile and Mac and Windows apps, and raised custom instructions to 5,000 characters — up from 1,500 — for Plus, Business, Enterprise and Education users. Its Codex coding tool also restored the full 272,000-token context window for the GPT-5.6 Sol, Terra and Luna variants.

Why it matters: The frontier headlines get the attention, but incremental workplace features are where AI actually reaches most businesses. Longer custom instructions and a dedicated work mode make it easier to encode how your organisation writes, formats and reasons — turning a general chatbot into something closer to a house assistant. It's a reminder that the practical value often comes from the plumbing, not the next model number. See OpenAI's newsroom for the running list of updates.

Quick hits

• Meta opens its Muse Spark API. Meta launched a US public preview of its first paid developer API for Muse Spark 1.1, with new accounts getting US$20 in free credits to start building. More
• DeepSeek reportedly designs its own chip. Reports this month say DeepSeek is building a custom inference chip to cut its reliance on Nvidia and Huawei — still early, with no named foundry or prototype, but a sign of China's push for domestic silicon. More
• Money keeps chasing applied AI. AI-agent startups drew around US$1.8 billion across a dozen-plus deals this month, led by legal-AI firm Harvey's US$200 million Series C at a US$2.1 billion valuation. More
• Big cloud doubles down on services. Microsoft and AWS both moved to put people, not just platforms, next to customers — with multi-billion-dollar programmes to embed AI engineers and integration teams inside client organisations. More

Looking ahead

Watch three dates. Moonshot is due to publish Kimi K3's full open weights by 27 July, which will test how quickly the open-source community can put a 2.8-trillion-parameter model to use. The EU AI Act's transparency obligations take effect on 2 August, and Australia's proposed AI standards head to National Cabinet the same month. Expect Google to follow its Gemini 3.5 Pro release with official benchmarks and pricing that either confirm or temper this week's early reporting.

That's the wrap for this week. The models are getting cheaper and the rules are getting firmer at the same time — a healthy sign that AI is maturing from spectacle into infrastructure. Check back next week for the next instalment of What's New in AI.