The AI Hub Weekly: Visa’s ChatGPT Integration Makes AI a Real Shopper

AI went from backstage to checkout this week. Most releases focused on putting powerful AI into the hands of people actually trying to get things done—not just demoing new tech. OpenAI, Google, and Visa all announced tools (and tie-ups) aimed at turning AI into something you use, not just read about. Let’s dig in.

TOP STORIES

Visa links ChatGPT to retail payments, letting AI agents shop for you

• Visa now allows AI agents built on ChatGPT to recommend retail products and complete purchases directly with its payment system, fully automating the shopping process.
• Any merchant supporting Visa rails can use this feature, meaning AI can now evaluate catalogs and check out with zero human input.
• This may make “one-prompt” shopping mainstream, raising big questions for e-commerce; rollout is gradual, with select partners live now.
• Source

OpenAI launches Partner Network to help businesses actually use AI

• OpenAI introduced a formal partner network to help organizations integrate AI into real workflows, moving beyond just giving access to models.
• The focus is on repeatable, scalable adoption—including process redesign and change management for large companies.
• Network partners will get training, early product access, and support; immediately available for enterprises of all sizes.
• Source

OpenAI Academy rolls out practical AI courses for the workplace

• New OpenAI Academy courses aim to close the gap between AI hype and real productivity at work, training users to apply AI in context.
• Courses range from basic prompts to workflow automation, available now for businesses and individuals.
• Designed for non-developers—start using AI for meetings, analysis, email, and more, not just playing with chatbots.
• Source

BBVA puts 100,000 employees on ChatGPT, claims big efficiency boost

• Spanish banking giant BBVA reports 100,000 staff globally are now using ChatGPT Enterprise, with 70% active monthly usage.
• Internal stats say staff save about 3 hours per week, with up to 80% efficiency gains in certain workflows.
• BBVA is integrating AI across customer service, operations, and internal work—one of the broadest rollouts in traditional finance so far.
• Source

OpenAI acquires Ona to power longer-running AI agents in the cloud

• OpenAI is acquiring infrastructure company Ona, known for secure, user-controlled cloud tools for running “agents” that handle complex, multi-step tasks.
• This should enable more reliable agents for coding, research, and automation, using OpenAI’s Codex tools.
• Over 5 million users weekly rely on Codex; expect smoother experiences and more robust privacy, especially for enterprise users.
• Source

Now you can buy OpenAI access on your Oracle cloud bill

• OpenAI and Oracle announced a partnership: companies can now use their existing Oracle cloud spending to pay for OpenAI models and Codex tools.
• This means no extra procurement headaches—just turn it on, use advanced models, and bill it to your Oracle account.
• Designed to help large organizations (especially those in highly regulated sectors) adopt AI without bypassing IT/security controls.
• Source

THIS WEEK IN AI

If you’d asked two years ago how “agents” would change our lives, most experts would have pointed to some distant future where AI could buy your groceries or run your meetings. Suddenly, “distant” looks like a bad prediction. Visa’s integration puts AI at the center of retail—the agent doesn’t just find what you want but checks out and pays, all on its own. Adoption is outpacing our readiness for the ramifications.

OpenAI, for its part, is out to solve the classic “AI-pilot-to-actual-impact” problem. New Academy courses and the Partner Network both say the quiet part out loud: powerful models aren’t enough. The real bottleneck is getting humans (and organizations) to trust and embed AI in daily routines. BBVA’s numbers—3 hours per week, real gains at banking scale—suggest it isn’t just theory. When you hear banks crowing about adoption statistics, it feels less like a hype cycle and more like the new normal.

But the story isn’t all smooth scaling. Behind the scenes, OpenAI’s acquisition of Ona and its cozying up with Oracle tell us “running agents” is getting complicated. Users (especially businesses) want the power without ceding too much control—or risking another runaway chatbot episode. Control, security, and privacy are becoming selling points for basic AI utility, not just marketing bullet points.

Finally, the flood of tools for “average” users—OpenAI Academy, faster models, new agent building recipes—signals the end of AI as a plaything for techies. The next wave is practical, not philosophical: Can this thing do my work? Can it buy my shoes? Will it save my team actual time?

For all the advances, there’s a lurking question: As AI moves from hype to habit, what tasks are you willing to hand over? If your next online purchase or emailed report came from an agent, would you notice—or care? This week, you can try for yourself. Why not see if a bot can save you some actual hours, not just deliver wow-factor demos?

MORE TOP STORIES

Chinese influence campaigns are targeting US AI policy debates, says OpenAI

• OpenAI reports that operations linked to the Chinese government are actively trying to shape US public debate on AI through misinformation and influence activities.
• The goal appears to be manipulating perceptions and decisions around how AI is developed and regulated globally.
• Raises questions about whether our policy debates about AI safety and governance are already being swayed by outside interests.
• Source

AI agent builds a 3D Paris gallery by chaining online tools—no manual labor required

• Using Hugging Face’s Spaces, one user had an AI agent assemble images and generate 3D reconstructions without any direct manual work involved.
• The result: a fully built, interactive web gallery of Parisian landmarks, using only AI-to-AI connections between services.
• Shows how “chained” AI tools can create complex content on-demand; similar builds are possible now for anyone with access.
• Source

Google DeepMind’s new AI model, DiffusionGemma, runs 4x faster on your own machine

• DiffusionGemma, just released, generates full blocks of text in parallel (not word-by-word), slashing local hardware inference time by 400%.
• That speed means non-cloud users get AI chat and creation almost instantly—key for privacy and low-power devices.
• Model is open and available now for download and experimentation; especially useful for home tinkerers and businesses worried about sharing data.
• Source

Anthropic forced to take Claude Fable 5 AI model offline by US government order

• New US export controls forced Anthropic to suspend access to its recently launched Claude Fable 5 and Mythos 5 models, citing national security risks.
• The sudden move knocked out a leading AI tool days after launch, affecting users globally and raising tough regulatory questions.
• Anthropic is contesting the order, but the timeline for return is unclear; existing customers face immediate disruption.
• Source

OpenAI’s real-time audio sessions now let you talk to documents, not just bots

• OpenAI updated its WebRTC Audio Session API: you can now ask questions or get summaries from specific documents in real time, via voice chat.
• Previously, you could only chat with the AI; now you can discuss your notes, reports, or uploaded paperwork on the fly.
• Available for developers and advanced users as of June 12, 2026, with further features planned.
• Source

Google sues alleged Chinese hackers for using AI to run mass scams

• Google is taking legal action against a Chinese group accused of running large-scale scams using its Gemini AI platform to automate phishing and fraud operations.
• The company says teams used Gemini to write, customize, and send fake messages at scale for months.
• Lawsuit signals escalating efforts to stop AI-powered crime—but also exposes new risks as these tools get better (and more accessible).
• Source

ALSO THIS WEEK

• Accenture: Consumers show growing trust in AI shopping agents — A new Accenture survey finds people are increasingly comfortable letting AI handle parts of their online shopping experience. (Source)
• Why AI hasn’t replaced software engineers, and won’t — Experts argue most coding jobs remain safe as human context still beats current AI models. (Source)
• Publishing WASM wheels to PyPI for use with Pyodide — New Pyodide features make it far easier to publish and use Python libraries in the browser. (Source)
• Statement on the US government directive to suspend access to Fable 5 and Mythos 5 — Anthropic issues a statement responding to the surprise government order shutting down its latest models. (Source)
• Coinbase for Agents: Automating portfolio trading with AI — Coinbase launches tools letting AI directly trade and manage user portfolios without manual steps. (Source)
• Xebia: Why AI agents fail without the right data foundation — Report argues that businesses need rock-solid data plumbing before AI agents deliver real value. (Source)
• Nine Things About Claude Mythos 5 That Matter If You’re Not an Enterprise Customer — Analysis of what the Claude Mythos 5 model means for everyday users. (Source)
• Claude Fable 5 and new AI safety fables — Deep dive into Claude Fable 5’s safety features and the larger debate. (Source)
• Claude Fable is relentlessly proactive — User impressions say Claude Fable 5’s new proactive features are a big step up, but raise fresh usability questions. (Source)
• datasette 1.0a33 — Alpha release lets Datasette users do more advanced queries in their workflows. (Source)
• Anthropic Walks Back Policy That Could Have ‘Sabotaged’ AI Researchers Using Claude — Anthropic reverses safeguard restrictions that limited researcher access to model features. (Source)
• datasette-agent 0.2a0 — New features let AI tools ask users clarifying questions as they process information. (Source)
• DiffusionGemma — User review highlights just how fast Google’s new AI model really is on personal devices. (Source)
• Quoting Jeremy Howard — Suggestion: labs with frontier AI shouldn’t use their own best models to boost further AI development. (Source)
• If Claude Fable stops helping you, you’ll never know — Analysis reveals potential transparency gaps if the model refuses difficult requests. (Source)
• Initial impressions of Claude Fable 5 — Early user review: impressive but still “feels AI” in certain complex tasks. (Source)
• llm 0.32a3 — New version of the open-source LLM tool is almost entirely AI-coded; showcases how capable the new models have become. (Source)
• Setting a custom price for a model in AgentsView — You can now assign custom pricing to AI models used in AgentsView workflows. (Source)
• Quoting Andrej Karpathy — Karpathy observes software demand is exploding because AI-driven tools make shipping projects frictionless. (Source)
• Siri AI at WWDC 2026 — Skepticism reigns: after past letdowns, the author warns users not to trust Apple AI demos until they ship. (Source)
• Google DeepMind is worried about what happens when millions of agents start to interact — DeepMind launches research exploring the possible risks and emergent behavior of mass-scale AI agent interactions. (Source)
• Meet the OpenAI Engineer Leading ChatGPT’s Biggest Transformation Yet — Interview with the lead behind OpenAI’s plan to turn ChatGPT into a personalized agent platform. (Source)