What's New in AI: The Power Shift — Week of 6 July 2026

The centre of gravity in AI shifted this week, and not in the direction most people expected a year ago. Anthropic shipped a cheaper, more capable Claude that runs agents at a fraction of last year's cost, quietly overtook OpenAI on revenue, and did it all while OpenAI was offering the US government a slice of the company. Closer to home, nearly every Australian government agency scrambled to appoint an AI officer by a 1 July deadline. Here is your plain-English roundup of what happened and why it matters for your business.

Anthropic's Claude Sonnet 5 makes cheap agents the new normal

On 30 June, Anthropic released Claude Sonnet 5, the most agentic version yet of its mid-sized model. The pitch is simple: it can plan, use tools like browsers and terminals, and run autonomously at a level that a few months ago needed much larger, pricier models. Anthropic says its performance sits close to the flagship Opus 4.8, and on one knowledge-work benchmark it edges slightly ahead. It is the default model for free and Pro users and is available on every plan.

The headline, though, is price. Sonnet 5 launched at US$2 per million input tokens and US$10 per million output tokens through 31 August, rising to US$3 and US$15 after that — cheaper than Opus 4.8, OpenAI's GPT-5.5 and Google's Gemini 3.1 Pro, as TechCrunch reported. Why it matters: agentic capability is no longer the differentiator — cost and reliability are. When a model can finish a two-step job like updating a CRM and sending a launch email without stalling halfway, the automation you shelved last year as too expensive suddenly pencils out. That is precisely the shift the team that builds and deploys custom AI for our clients has been preparing for: the economics of putting AI to real work have quietly flipped.

OpenAI offers Washington a 5% stake as frontier rules take shape

In one of the stranger stories of the week, Sam Altman floated giving the US government roughly 5% of OpenAI — a stake worth about US$42.6 billion — as donated equity that would channel AI profits into a sovereign-wealth-style vehicle rather than requiring any taxpayer cash outlay. The idea, first pitched in early 2025, resurfaced in Washington this week alongside broader talks about how frontier models should be governed. Notably, Anthropic is not participating in the equity discussions.

It lands as the White House moves toward voluntary release standards for the most advanced models, in advanced talks with OpenAI, Anthropic and Google to set security benchmarks, review timelines and rules on who can access frontier systems at home and abroad. It is the natural next step after the export-control drama we tracked in last week's edition. Why it matters: when the government holds equity in a company it also regulates, the conflict-of-interest questions write themselves, and several experts have already raised them. For businesses, the takeaway is that access to the very top tier of models is becoming a policy question, not just a pricing one — another reason to build on the broadly available models you can actually count on.

Anthropic quietly overtakes OpenAI on revenue

The rivalry tipped over a milestone this week: by several accounts Anthropic's annualised run-rate revenue has moved ahead of OpenAI's. Reported figures vary — Anthropic has been described as crossing a run-rate north of US$30 billion, with some reports putting it closer to US$47 billion, against OpenAI's roughly US$24–25 billion — but the direction is consistent across sources, and it arrived earlier than forecasters like Epoch AI had predicted. It follows Anthropic overtaking OpenAI as the most valuable AI startup in late May.

The gap comes down to strategy. Anthropic has leaned hard into enterprise and developer customers, while OpenAI built its business on consumer ChatGPT subscriptions — and ChatGPT's share of generative-AI usage slipped below a majority for the first time earlier this year. Why it matters: the "default" AI vendor is no longer a settled question. For businesses choosing a platform to standardise on, a genuinely competitive top two means better pricing, faster feature releases and less lock-in risk than the one-horse race many assumed was coming.

Australia's public service races to appoint AI officers

Australia had its own AI deadline this week. Under the Commonwealth's push to lift public-sector AI maturity, agencies were required to appoint a Chief AI Officer by 1 July — and nearly all made it, with 104 of 106 non-corporate Commonwealth entities appointing one, according to Information Age. The number of agencies with a CAIO roughly doubled in the final fortnight as the deadline loomed.

It sits within the broader National AI Plan, which promises every public servant access to generative-AI tools plus training, and funds a new AI Safety Institute with about AU$29.9 million. Rather than sweeping new laws, Canberra is building on existing regulatory frameworks. Why it matters: government is often the slowest mover, so when 104 agencies name an AI lead in a fortnight, it signals that "someone owns AI here" is becoming a baseline expectation. If you run a business that sells to, partners with, or competes against government-adjacent organisations, expect AI literacy and clear accountability to become table stakes in procurement and tenders.

Google and the coding-tool makers keep shipping

The product treadmill did not slow down. On 30 June, Google released two new image models — Gemini 3.1 Flash Image and Gemini 3 Pro Image — both live immediately through Google AI Studio and the Gemini API, and rounded out a busy June that included Live Translate and a beefed-up NotebookLM that now runs code and generates charts, spreadsheets and slide decks. Meanwhile Cursor launched a native iOS app in public beta, letting developers kick off and steer cloud coding agents from their phones, complete with voice input and lock-screen live activities.

The common thread is control: these tools increasingly let you dispatch and supervise AI work rather than just chat with it. Why it matters: the practical frontier for most businesses is not the smartest model but the best-integrated workflow — the ability to point AI at a real task and check its output. If you are trying to keep track of which tools are worth your time, our running library of AI tools and resources is a good place to start. The winners this year will be teams that wire these agents into everyday operations, not those chasing the newest benchmark.

The money keeps pouring in — and it is chasing compute

Investment showed no sign of cooling. Early July brought Together AI's US$800 million Series C, reportedly led by Aramco Ventures, while infrastructure builder Crusoe was said to be in talks to raise around US$3 billion and energy-focused Joulent secured US$1.75 billion in strategic financing. The pattern is telling: much of the biggest money is now flowing not into apps but into the data centres and power that AI runs on.

AI has captured close to half of all global venture funding, and by one tally nearly 88% of AI startup money this year has gone to US-headquartered companies. Why it matters: when capital concentrates on compute and energy, it is a signal that demand for AI is expected to keep climbing for years, not quarters. But the geographic tilt is a caution for Australian founders and operators — the funding boom is real, yet it is not evenly shared, so local businesses may need to lean on capability and customer proximity rather than out-raising overseas rivals.

Quick hits

• Anthropic's flagship Fable 5 returned to users worldwide on 1 July after the US Department of Commerce lifted the export controls it had imposed in June — closing out a month-long saga.
• Salesforce agreed to buy AI customer-service platform Fin for about US$3.6 billion, deepening the race to automate support. (details)
• ON Semiconductor struck a roughly US$7 billion all-stock deal for Synaptics to push into "physical AI," per CNBC.
• China's Meituan open-sourced LongCat-2.0 under a permissive MIT licence, unmasking it as "Owl Alpha," the anonymous model that had topped OpenRouter's developer charts for weeks, as VentureBeat reported.
• Google's Gemini 3.5 Pro was cleared for general availability in July after slipping from a June target.

Looking ahead

Watch for the White House to formalise its voluntary frontier-model standards, which reporting suggests could land within days — the details will shape who gets access to the most powerful models. Keep an eye, too, on how quickly the cheaper agentic tier (Sonnet 5 and its rivals) shows up inside everyday business software, and whether Anthropic's revenue lead holds once OpenAI's GPT-5.6 family reaches wider availability.

That is the wrap for this week. The through-line is hard to miss: AI is getting cheaper to run, harder to separate from government, and more competitive at the top — all of which is good news if you would rather adopt the technology than bet on a single winner. Check back next week for the next instalment of what's new in AI.