What's New in AI: Opus 5 Lands, Rules Tighten — 27 Jul

The models kept getting cheaper this week while the rulebook kept getting heavier. Anthropic dropped a surprise new flagship, Google shipped a faster and cheaper Gemini, and Microsoft tightened its grip on enterprise AI — even as Australia and the European Union moved to make AI companies pay for the power they burn and the creative work they train on. If there's a single through-line, it's this: the cost of using capable AI is falling fast, while the cost of using it carelessly is starting to rise. Here's your plain-English roundup of what happened and why it matters for your business.

Anthropic's Claude Opus 5 undercuts its own frontier

On Friday 24 July, Anthropic launched Claude Opus 5, the newest version of its heavyweight model — just two months after Opus 4.8. Although smaller than the flagship Fable 5, Opus 5 is cheaper, less restricted, and actually beats Fable 5 on several of the benchmarks Anthropic published. The company says the model is “much stronger at verifying its work and iterating carefully until it succeeds.” A new opt-in feature called Automatic Fallbacks quietly reroutes a request to a smaller model when a safety filter trips, so developers get a working answer instead of an error, and Anthropic expects its safety classifiers to fire roughly 85% less often than they do on Fable 5.

Why it matters: for businesses, this is the clearest sign yet that near-top-tier reasoning is becoming a commodity. When a cheaper model outperforms the flagship on real tasks, the question shifts from “can we afford frontier AI?” to “which capable model actually fits the job?” That is precisely the calculation our AI development team helps clients work through when they move a promising prototype into day-to-day production.

Google ships a cheaper Gemini — and starts training Gemini 4

A day earlier, on 21 July, Google released a trio of models led by Gemini 3.6 Flash. It uses about 17% fewer output tokens than the version launched at I/O in May, finishes multi-step jobs in fewer steps, and costs less — US$1.50 per million input tokens and US$7.50 per million output. Its knowledge now runs to March 2026. Alongside it came the lightweight 3.5 Flash-Lite for high-volume work and a security-tuned 3.5 Flash Cyber built to find and patch code vulnerabilities. Google also confirmed its long-delayed 3.5 Pro is still in partner testing, and that it has begun its “most ambitious pre-training run yet” for Gemini 4.

Why it matters: Flash-class models are what most businesses actually run in production — powering chatbots, summarising documents and processing paperwork — because they are fast and inexpensive. A 17% cut in token use is a direct cut to your monthly AI bill. If you're weighing which model to build on, it pays to compare the current crop side by side, which is one reason we keep a running view of the major tools and models in one place.

Microsoft tightens its grip on enterprise AI

While the model makers grabbed headlines, Microsoft spent the week locking down the enterprise layer. On 23 July it extended its decade-long partnership with data platform Databricks into the 2030s, pushing Databricks' “Genie” AI co-worker directly into Microsoft workflows and moving Databricks' own operations onto Azure. Two days earlier it deepened its tie-up with France's Mistral, bringing Mistral's models to Azure and Microsoft Foundry with a focus on “sovereign” cloud for regulated industries.

Why it matters: the frontier models get the attention, but the real enterprise battle is over whose platform grounds AI in your company's own data — your customers, products, metrics and processes. Microsoft is betting that owning that connective tissue matters more than owning the single smartest model. For business owners, the takeaway is that the value of AI comes less from the model itself and more from wiring it reliably into your actual data and workflows. A brilliant model that can't see your customers, stock or paperwork is just an expensive demo.

Australia moves to make AI pay for power — and for creators

Prime Minister Anthony Albanese signalled some of the world's toughest AI rules, built around mandatory national standards due to be legislated in early 2027. Two proposals stand out. New data centres would have to become net energy producers — feeding at least as much power into the grid as they draw, funding their own connections and meeting water-efficiency standards. And Australia would reject the broad “text-and-data-mining” exemptions used elsewhere, so writers, musicians, artists and journalists keep control of their work and must consent to, and be paid for, its use in AI training. Albanese called anything less “theft.” Canberra estimates AI could add about A$116 billion to the economy over the next decade. It builds on the national AI office we covered in last week's roundup.

Why it matters: this is the local story with the longest tail. If you run a data-heavy business, or you own creative work — copy, photography, music, journalism — the ground rules for how AI can use Australian energy and content are being drawn right now. They also hint at where compliance costs, and new licensing opportunities, are likely to land.

The EU's big AI deadline is days away

The European Union's AI Act hits a major milestone on 2 August 2026. From that date the European Commission can enforce its rules for general-purpose AI models — with fines reaching up to €15 million or 3% of global turnover — and the compliance framework for many “high-risk” AI systems comes into application. Providers' obligations for general-purpose models have technically applied since August 2025, but 2 August is when the power to police them, and to fine, switches on.

Why it matters: the AI Act reaches any business that offers AI-touched products or services to EU users, not just European firms. Even Australian companies selling into Europe may need to document how their AI systems work, classify their risk level and name someone responsible for them. If Europe is in your market, the next fortnight is a sensible time to check where your AI use sits.

Quick hits

• Anthropic's $1.5 billion settlement with authors over pirated training books received final court approval on 20 July — the first major AI copyright case in the US to settle, paying roughly $3,000 per work.
• SAP completed its €1 billion-plus acquisition of Germany's Prior Labs, betting on “tabular” AI models built for the structured data inside spreadsheets and databases rather than free-form text.
• Together AI raised $800 million at an $8.3 billion valuation to scale cheaper open-source model hosting, underlining how much money is now chasing lower-cost inference.
• Google is reportedly developing a new in-house AI chip to make running Gemini cheaper and more efficient — part of an industry-wide push to drive down the cost of serving models.

Looking ahead

All eyes turn to 2 August, when the EU's enforcement powers switch on and compliance stops being optional. Watch too for Google's much-delayed Gemini 3.5 Pro, and for any pricing answer from OpenAI to Opus 5 — the race to make capable AI cheaper shows no sign of slowing.

That's the week in AI. The pattern is hard to miss: capability keeps getting cheaper while regulation keeps getting real, and both trends reward businesses that are deliberate about how they adopt these tools. We'll be back next week with the stories that matter — see you then.