ChatGPT Just Got a Visa Card
For two years, AI has answered questions. This week, it started paying for them.
“On 10 June at the Visa Payments Forum, Visa and OpenAI announced a partnership embedding Visa's payment network directly inside ChatGPT — letting AI agents complete purchases at any Visa-accepting merchant under user-set spending limits and authorisation controls. On 9 June, Anthropic released Claude Fable 5, a Mythos-class frontier model now generally available on the Claude API, AWS Bedrock, Vertex AI and Microsoft Foundry, with a 1-million-token context window and a 10-point benchmark jump over the prior Opus. And on 11 June, OpenAI agreed to acquire Ona, the startup behind the secure, pre-configured cloud environments that let AI agents complete long-running tasks — folding it into the Codex group, which now serves 5 million weekly users, up 400% this year.”
Three announcements in five working days: the components needed to let software act on an owner's behalf — a frontier model, a place to run it, and a way to pay — all shipped from different vendors inside the same week.
Here's what happened, why it matters, and what your business should do about it.
The Big Three
1. Visa Plugged Itself Into ChatGPT
At the Visa Payments Forum in San Francisco on 10 June, Visa and OpenAI announced that Visa's payment rails will sit directly inside ChatGPT. Once a user links a Visa card, a ChatGPT agent can browse, evaluate and complete purchases at any Visa-accepting merchant — with Visa managing authorisation and fraud monitoring, and the user retaining permission gates, spending limits, and required approvals. The two companies also said they will collaborate on enterprise applications using OpenAI's Codex assistant.
The signal is not "people now shop in a chatbot." It is that the buy-side of e-commerce is being rewritten. A Malaysian seller on Shopee, Lazada, TikTok Shop or a Shopify storefront does not get to opt out of how their international customers' AI agents discover, evaluate and pay them. If your product page, pricing structure and review profile are not machine-readable today, you are about to lose impressions to competitors whose are.
2. Anthropic Released Claude Fable 5 — Mythos-Class for the Public
On 9 June, Anthropic released Claude Fable 5, described as a Mythos-class model "made safe for general use." It is generally available on the Claude API, AWS Bedrock, Vertex AI and Microsoft Foundry, with a 1-million-token context window, up to 128K output tokens, and the first 90%+ score on Anthropic's core analytics benchmark — a 10-point jump over Opus. API pricing is US$10 per million input tokens, US$50 per million output, with a 90% discount on cached prompt content.
The dull, useful reading for an SME owner: frontier reasoning capability got materially better this week, and it now lives in exactly the cloud platforms (AWS, Microsoft Azure, Google Cloud) your IT vendor already buys from. "We can't get a frontier model into our stack" stops being a credible objection. The question is whether the work you would point it at is itemised anywhere.
3. OpenAI Bought the Sandbox the Agents Run In
On 11 June, OpenAI announced an agreement to acquire Ona, a startup that builds secure, pre-configured cloud environments for AI agents — isolated sandboxes giving agents the tools, files and context to complete long-running tasks reliably. Financial terms were not disclosed; the team folds into the Codex group, where weekly active users have reached 5 million, up 400% this year.
A fortnight ago, Microsoft's Autopilots framework solved whose identity an agent acts under. This week, OpenAI bought the answer to where the agent does its work. The platforms are systematically buying or building every component an enterprise needs before agents can be production-grade. The closer that completeness gets, the less of an experiment "deploying an agent" becomes — and the more it looks like normal procurement.
Closer to Home: Malaysia
The Malaysian story this week sits at Sunway University today: the National AI Competition 2026 (NAIC) grand finals, themed around the UN Sustainable Development Goals, with RM2 million in bursaries awarded across five tracks. Competitors are pre-university students aged 14 to 19 — not enterprise engineers — but the signal is the pipeline, not the prize pool. Pair it with Budget 2026's 50% additional tax deduction on certified AI and cybersecurity training for SMEs, and locally-trained AI talent is being subsidised into the labour market the same year their future employers are subsidised to hire them.
While that runs in Subang, the e-invoice clock keeps ticking. Phase 4 businesses (RM1–5 million turnover) still have the grace period to 31 December 2027, with RM200–RM20,000 per-invoice penalties from 1 January 2028. The RM10,000-and-above individual e-invoice rule is live now, and February's LHDN figures — 500,000+ non-compliant cases, RM1.4 billion in unreported income — show audit risk is not theoretical during the grace window. In the week global headlines were about agents paying for things on your customers' behalf, the local headline is that the audit trail on what you charged them must already be machine-issued.
Take the cross-border layer seriously too. The Visa–OpenAI deal makes Malaysian sellers — particularly those catching Singaporean, Australian and Gulf buyers through Shopee, Lazada, TikTok Shop or Shopify — visible to a buying channel that did not exist a fortnight ago. The seller-side prep is unglamorous: a clean product feed, structured attributes, complete pricing in major currencies, machine-readable shipping policies, and review hygiene. None of it is exotic. All of it is overdue.
Singapore, briefly: the agentic-commerce stack is widely expected to pilot through Singapore and selected ASEAN markets first, consistent with Mastercard's earlier ASEAN agentic-payments rollout. Expect AI-driven buying behaviour to land in Singapore ahead of the local market.
What This Means for Your Business
1. Audit your product data the way an agent would read it
Open your top ten product listings on every channel you sell on — Shopee, Lazada, TikTok Shop, Shopify, your own site. Score each one on title clarity, attribute completeness, pricing structure, shipping detail, return policy, and review volume. If you would not buy it from the listing yourself in 30 seconds, an agent comparing fifty similar listings in two seconds will not surface it either. The Visa is in the agent's wallet now; the question is whether your page lets it click buy.
2. Move your highest-volume RM10,000+ transactions onto individual e-invoices this month
Not in October. This month. The rule is live, the penalty regime starts 1 January 2028, and the implementation queue at the bottom tier of consultants gets ugly in late 2027. The matching 50% MADANI digitalisation grant up to RM5,000 still covers half the tool cost on MDEC-listed providers. Treat the runway you have been given as a quality buffer, not a procrastination buffer.
3. Pick the one annual document your team still assembles by hand
Fable 5 is overkill for an email reply. It is appropriately sized for the year-end customer report, the quarterly board pack, the tender submission you piece together from five other documents. Price the largest such task in your business against the API rate (a 50,000-word draft runs roughly US$20 in tokens) and decide whether the four-figure ringgit of staff time it replaces is worth keeping in-house. The arithmetic was unfavourable in January. It is not today.
The Practical Question
If a customer's AI agent visited your business this week — could it figure out what to buy, at what price, on what terms, and pay you — without a human in the loop on your side?
The week's quiet pattern is that every layer of agentic commerce now has a vendor. Visa carries the money. Anthropic carries the reasoning. OpenAI carries the sandbox. Microsoft carries the identity. Apple carries the device. The three major clouds carry the runtime. None of it asks your business to integrate anything new. All of it asks your business to be findable and buyable on a clean machine-readable contract.
The owners who win the next year will treat agentic commerce as a listings audit, not a strategy project.
At The Empyrean, we help Malaysian SMEs find the practical, repeatable tasks where AI delivers value without disruption. If you're not sure where to start, we're happy to take a look at your operations and tell you honestly what would make sense.