Apple has announced its third-generation Apple Foundation Models — AFM 3 — a family of five AI models ranging from what runs inside the phone to what lives on the servers. The detail that drew most attention: they were built in partnership with Google.

For anyone who has followed the saga, it makes sense. Apple arrived late and awkwardly to the AI race, with a Siri that promised much and delivered little. Rather than trying to do everything alone, the Cupertino company seems to have decided it is better to lean on someone who already has muscle in the field.

On-device and in the cloud

The logic of AFM 3 is to split tasks: smaller models run directly on the device, fast and more privacy-friendly because the data never leaves the phone; larger ones sit on servers for heavier requests. It is the same balance the whole industry is chasing, between speed, cost and data protection.

The big question is whether this is enough for Apple to claw back ground from OpenAI and Google within its own products. Having good models is one thing; turning them into an assistant people actually use every day is quite another — and that is precisely where Apple has stumbled.

See also: Apple and the choice of an AI partner. Official announcements are at apple.com.

Gemini Omni Flash is now in developers’ hands. Since the end of June, Google’s video model has been available through the Gemini API and Google AI Studio, with pricing that’s easy to memorise: $0.10 per second of generated video. After May’s consumer debut at Google I/O, this is the step that takes Omni out of Google’s own app and into everyone else’s products.

How much does Gemini Omni Flash cost on the API?

Ten cents per second of output — a ten-second clip runs about a dollar. The model generates and edits video conversationally: hand it a photo, a text prompt, an audio clip or another video, describe the change in plain words, and the result comes back in seconds. If you’d rather not write code, Omni Flash remains available in the Gemini app for AI Plus, Pro and Ultra subscribers, and free of charge on YouTube Shorts.

What does it change for creators and businesses?

API access means any application — a marketing tool, a video editor, an e-commerce platform in Portugal — can generate made-to-measure video without building a model of its own. With Gemini now past 900 million monthly users across 230 countries, Google is turning AI video into a commodity, priced at roughly a coffee per minute of footage. Competitors may want to hold on to the script.

The usual question remains: the cheaper and easier realistic video becomes to fabricate, the more valuable the ability to tell real from synthetic. Europe’s regulatory answer is still being written.

See also: the Gemini Omni launch and what it can do. Technical details and pricing are on Google’s official blog and Google AI Studio.

Artificial intelligence finally has a table where everyone gets a seat. The Global Dialogue on AI Governance opens today, 6 July, in Geneva, bringing together the UN’s 193 member states, tech companies, academia and civil society for two days — the first time AI governance is being discussed on a universal platform rather than in restricted clubs of wealthy countries.

What is the Global Dialogue on AI Governance?

It is the forum created by the UN General Assembly under the Global Digital Compact, for governments to share practices and converge on approaches to regulating artificial intelligence. It produces no binding rules — the model is the Internet Governance Forum: two days of discussion and a co-chairs’ summary at the end. This first session runs on 6-7 July in Geneva; the second is scheduled for New York in May 2027.

Is it useful if it binds no one?

That is the right question — and the honest answer is: it depends what you ask of it. Nobody leaves Geneva with a “world AI law”. But at a moment when the technology is evolving faster than any national regulator can track, and warnings about serious risks keep piling up, having 193 countries in the same room is the first step towards rules that are not written by half a dozen capitals and half a dozen companies. For Europe — and Portugal — betting on its own path for AI, it is also a stage to defend that vision.

See also: Europe’s push for sovereign AI. Official programme at un.org.

End of the road for Google. The Court of Justice of the European Union, the bloc’s highest court, threw out the company’s final appeal and confirmed the fine of around €4.1 billion in the Android case. The ruling is final: there is nowhere left to appeal.

Why was Google fined?

The European Commission accused Google of abusing Android’s dominance to favour its own apps, pushing phone makers to pre-install Google Search and Chrome in exchange for access to the Play store. The original 2018 penalty was €4.34 billion; a lower court trimmed it to €4.1 billion in 2022, and that is the figure now set in stone.

What changes for Android users?

On your phone screen, little for now. But the judgment opens the door to damages claims from rivals who felt squeezed for years, and it strengthens Brussels’ hand in its other cases against Big Tech. For a country like Portugal, where the vast majority of phones run Android, it is one more sign that Europe wants clear rules on who controls the gateway to the internet in everyone’s pocket.

It was an eight-year courtroom battle, and Google lost across the board.

See also: Mistral and Europe’s sovereign AI push.

Official details at the Court of Justice of the EU.

Microsoft has decided that selling AI software by licence is no longer enough: it wants to walk into companies and build the systems itself. To do that it has created the Microsoft Frontier Company, a new unit with an announced investment of $2.5 billion.

What is Microsoft Frontier?

It is a team of roughly 6,000 people, spanning engineers, industry specialists and salespeople, who will work side by side with large clients to design, deploy and fine-tune bespoke artificial-intelligence systems. Instead of handing over a tool and wishing you luck, Microsoft will now sit its own experts inside the customer’s building.

What changes for companies?

It changes how AI actually enters a company: instead of buying a licence and waiting for results, the client now gets Microsoft’s own people building and tuning the systems from the inside. It confirms that the AI battle is no longer just about the models, but about deployment — and whoever helps make it run day to day locks in that customer for years. For Portugal’s business fabric, increasingly reliant on these platforms, it is a reminder that tech dependence is also measured in who does the installing.

The move shows how fast the giants want to turn promise into hard revenue.

See also: Anthropic’s bet on science.

More at Microsoft’s official newsroom.

Sovereign AI in Europe is the continent’s ability to train and run its own artificial intelligence — in its own data centres and under its own rules — without depending on companies from elsewhere. It is the bet growing much closer to home while the world talks about OpenAI, Google and Anthropic, almost all American, and France’s Mistral has become its face.

What is sovereign AI in Europe?

It is artificial intelligence trained and operated inside Europe, on European infrastructure and under European law. For the bloc, it is a matter of strategic autonomy: sensitive data that stays home, models that understand European languages, and servers that are not hostage to decisions taken on another continent.

Why is Mistral leading this race?

Mistral, founded in Paris, has been named Europe’s champion for sovereign computing. It joined Nvidia’s Nemotron coalition as a founding member, to co-develop open models, and secured 830 million dollars for an Nvidia-powered data centre near Paris, due to come online in 2026. In parallel, Nvidia has pledged to help stand up more than 20 “AI factories” across the continent — facilities designed to produce artificial intelligence at industrial scale.

Why does this matter to Portugal?

Portugal sits on the European side of this board: the closer the AI infrastructure is, and the better the models handle Portuguese, the less dependent we are on outside platforms. It is the difference between using other people’s technology and having a say in how it is built.

See also: Europe’s new rules for labelling AI content. More about the company on Mistral’s official site.

If you thought artificial intelligence was now just for writing emails and summarising meetings, early July brings a different message: the big companies want to put AI to work doing serious science.

Claude heads to the lab

Anthropic announced an app dedicated to scientific research, with the stated ambition of shortening research and development cycles in the life sciences — read: helping to discover drugs and test hypotheses faster. The move comes with high-profile hires and an acquisition in the hundreds of millions of dollars, a sign this is more than a marketing announcement.

Behind it is an idea that is simple to explain and hard to pull off: if a model can read thousands of papers, suggest experiments and analyse results, it might compress years of lab work into months. It is a bold promise, and it deserves the healthy skepticism owed to anyone who has heard bold promises before.

The flagship models in the waiting room

Meanwhile, the heavyweights are juggling calendars. Google pushed back its next flagship after enterprise customers complained it burned through too many tokens on long tasks. OpenAI held off on the wide release of its new generation at the request of the US government, which wanted to review the model before it reached the public.

In plain terms: the race is still intense, but it is no longer only about who ships first — it is about who ships reliably, without raising red flags with regulators.

For context, see what we wrote about Claude arriving on the iPhone and about the real cost of using AI. The official details are on Anthropic’s site.

OpenAI has shuffled the board again. The company has begun a limited preview of GPT-5.6, the next generation of its models, presented in three flavours with names that taste of the solar system: Sol, Terra and Luna. The promise is a jump in reasoning, coding, biology and even cybersecurity, plus a new ultra mode for the most demanding tasks.

For now, access is tight. During the preview, GPT-5.6 is available through the API and Codex only to a select group of trusted partners and organisations, with arrival in ChatGPT promised for later. There are also updated pricing rules and a more predictable prompt-caching system — technical details that, in practice, decide how much it costs to put these tools to work.

It is not just the model

In the same breath, OpenAI introduced GeneBench-Pro, a research-level benchmark for judging AI agents in computational biology, with harder and more realistic tasks. It is one more sign of where the race is heading: models that do not just chat, but try to solve concrete scientific problems.

For anyone living in Portugal and working with these tools, the pattern holds: the new features reach paying API users first and only later trickle down to the everyday user. It is worth not chasing every announcement and letting the dust settle. We had already covered OpenAI’s push into hardware and Codex, and this preview fits that strategy of occupying ground on every front. Official details are being posted on OpenAI’s site.

For years the iPhone had a single voice: Siri, for better and worse. At WWDC 2026, Apple rewrote the script and began letting users choose which artificial-intelligence assistant to use on the phone — and Anthropic’s Claude joined that list of options. For anyone with an iPhone in their pocket, it’s a small big deal.

What changes day to day

The idea is simple: instead of being stuck with one assistant, you can route more complex requests — writing, summarising, coding, organising — to the model you prefer. Apple keeps Siri as the gateway and the privacy layer, but hands the heavy reasoning to outside partners when it makes sense.

Why it matters

Two years ago the conversation was whether Apple had fallen behind in the AI race. By opening the iPhone to several models, the company turned a weakness into a position of strength: rather than trying to win alone, it became the stage where the best assistants compete for the user’s attention. For Anthropic and its rivals, reaching hundreds of millions of iPhones is the kind of shop window no marketing budget can buy.

How the bill shakes out remains to be seen — who pays what, and which data leaves the phone. But the direction is clear: the assistant on your phone has gone from an imposition to a choice.

See also: how Siri started using extensions and outside models. Official details and dates appear on Apple’s events page.

The holidays are here and, with them, the old question: can I use my phone freely abroad? The answer depends a lot on where you’re going. Inside the European Union, roaming has worked for years under the roam like at home rule — use data, calls and messages at the same price as in Portugal, with no hidden extras. You travel and carry on your digital life as if you were home.

Outside the EU, watch the bill

The trouble starts when you leave the European area. In many destinations, roaming data costs small fortunes per megabyte, and it only takes your phone updating apps in the background for the bill to spike without you noticing. The golden rule is simple: before travelling, check the conditions for your destination country with your operator.

Tricks to avoid overspending

There are easy fixes. Buy a temporary roaming pack from your operator, use a local eSIM for the destination, or simply switch off mobile data and lean on the wi-fi of hotels and cafés. Turning off automatic updates and photo backup while you’re away also saves a lot. That way you enjoy the maps and messages without the shock when you get back.

Connected is good; connected and relaxed is better.

See also: the iPhone opens up to new AI assistants. European roaming rules at the European Commission.

Anyone hoping to get their hands on the new Gemini 3.5 Pro will need a little more patience. Google confirmed it is pushing the launch of its latest artificial-intelligence model to July, after gathering feedback from testers.

Why the delay

According to the company, the delay is meant to improve two points that more technical users value: token efficiency — spending fewer resources to do the same work — and performance on long, agentic tasks, the ones where the AI has to chain many steps without losing its way.

It may sound like a detail, but this is where the next phase of the AI race is being played. Answering a question well is no longer enough: models have to be able to carry out complex tasks autonomously and reliably. And there is still plenty to fine-tune there.

What it means for us

For the average user, a delay like this is mostly a sign of how fierce the competition is. Google, OpenAI and company are shipping versions at a dizzying pace, and nobody wants to risk putting out a half-baked product.

Interestingly, this race is also fought over people: OpenAI has been poaching Google heavyweights, as we reported in the case of the Transformer’s creator. To follow the announcements first-hand, the best bet is Google’s official blog.

Artificial intelligence has lived almost entirely inside the screen. OpenAI wants to change that: the company is preparing to unveil, on July 15, new physical hardware for Codex, its coding agent.

A control panel for AI

The idea, from what is known, is to create a kind of physical control panel — a dedicated device that acts as a bridge between the developer and the AI agent writing code. Instead of being just another window on the computer, Codex would gain a presence of its own on the desk.

It fits OpenAI’s strategy. As AI agents become more capable of carrying out tasks on their own, it makes sense to give them a dedicated interface, designed to monitor and control what the machine is up to. It is a curious step: AI leaving pure software behind and starting to take physical form.

The money race

Behind these announcements is a torrent of investment that is hard to picture. Alphabet, Google’s parent, recently closed the largest equity financing in corporate history — around 84.75 billion dollars — just to feed its AI infrastructure. That is the scale at which the game is now played.

For Portugal, all this still feels distant, but it defines the tools we will be using tomorrow. See also our piece on the Transformer’s creator switching camps and follow the official announcements at openai.com.

Some names mean nothing to the general public but make AI engineers raise an eyebrow. Noam Shazeer is one of them, and news of his move is among the most talked-about in the industry this year.

On 18 June, Shazeer announced he’s leaving Google DeepMind to join OpenAI. Why does it matter? Because he co-authored a 2017 paper called “Attention Is All You Need,” the text that introduced the Transformer architecture — the piece of engineering underneath practically every AI model we use today, from ChatGPT to the translator on your phone.

A heavyweight transfer

In football you’d call this a marquee signing, and that’s no exaggeration. It’s the biggest individual talent move in AI since Andrej Karpathy changed houses, and it neatly illustrates the war for brains being fought among the big tech firms. Companies no longer compete only for chips and data centres; they compete for the handful of people who can design the next generation of models.

For OpenAI, closing out a phase of rapid growth, it’s both a symbolic and a practical boost. For Google, it’s a painful loss — all the more so because it’s someone who helped invent the very technology everyone is now mining.

What’s left unsaid

The real question is what Shazeer will build next. People who laid the foundations of a technology rarely change houses to do more of the same — and that’s what makes this hire so intriguing.

See also: Portugal’s startups and innovation support. More on the company at OpenAI’s official site.

How many times have you handed your number to a stranger just to sort out an online-marketplace sale or message a shop? WhatsApp wants to end that awkwardness. Meta’s app will start letting people talk through a username, without sharing a phone number — and name reservations are already open.

How it will work

The idea is simple: instead of swapping contacts, you share a unique username, Instagram- or X-style. To pick one, you go to Settings, Account and Username. It can be 3 to 35 characters, using lowercase letters, numbers, the _ symbol and periods. Anyone wanting to lock in their preferred name before someone else grabs it can reserve it now.

There’s also an extra layer of security: a username key, a code the other person needs to know, on top of the name, before they can message you. That stops just anyone writing to you by guessing the name.

What doesn’t change

Best not to get carried away: an account still needs a phone number to exist. And numbers you’ve already shared with contacts, or that show up in group chats, stay visible to whoever already has them. The change mainly protects those first contacts — with shops, marketplaces, communities or people you’ve just met.

For a country where WhatsApp is practically everyone’s phone line, it’s a welcome shift. Fewer numbers passed hand to hand is, almost always, more peace of mind.

See also: the new Siri and Gemini delayed in Europe. The official announcement is on the WhatsApp blog.

That nagging question of “is this real or made by AI?” is about to get a mandatory answer in Europe. The European Union’s Artificial Intelligence Act brings transparency rules that require content generated or manipulated by AI to be marked, so that machines and, ideally, people can tell it is not authentic.

What changes in practice

Images, videos, audio and text produced by AI systems will have to include a kind of digital watermark, detectable automatically. The aim is to curb disinformation and deepfakes, giving the user a clue that what they are seeing may not be a real photograph or a true statement by someone.

When it kicks in

The timetable has been adjusted. The transparency rules were due in August 2026, but the obligation to mark artificial content in a machine-readable way has been pushed to December 2026, giving companies more time to adapt. The Commission has also already appointed the scientific panel that will support enforcement.

For the average user, the promise is simple: fewer deceptions. It does not solve everything, there will always be those who try to dodge it, but it is a step toward an internet that can once again tell the genuine from the fabricated.

See also: Google creating video out of anything and Europe squeezing the giants.

Official text and timeline of the AI Act at digital-strategy.ec.europa.eu.