OpenAI and Microsoft hold dueling enterprise AI events on the same day — Codex repositioned as the interface for all knowledge work, Microsoft's seven new in-house models pitched on cost rather than benchmarks — while in the headlines, Trump's strange, once-pulled AI executive order finally gets signed in private with zero fanfare.
Enterprise AI's next wave runs on two tracks: fluency and cost.
As AI moves from the subsidy era to the token-shortage era, adoption is no longer just a question of what's possible — it's a question of tool and use-case fluency, and of cost. OpenAI's answer is Codex as the new interface for knowledge work; Microsoft's is near-frontier models tuned to your tasks at a tenth of the cost. If you want to simplify it: the second half of 2026 is going to be about wrestling into a workable, cost-effective approach all of the opportunities the first half unlocked.
Both point at the same shift: from the subsidy era to the token-shortage era, where agentic workloads push token consumption against the limits of available compute and business models realign. Enterprise adoption is now two problems at once — cost, and tool and use-case fluency.
OpenAI's "Next Era of Knowledge Work" report puts Codex at 5 million weekly active users — with non-technical knowledge workers adopting it three times faster than developers. Google searches for Codex passed Claude Code for the first time in May, and The Information called the vibe shift "palpable."
Workers can produce documents, dashboards, and presentations faster than ever, yet a McKinsey study OpenAI cites finds the average knowledge worker spends more than a quarter of the workweek managing email and almost a fifth hunting for internal information or the right colleague. Three frictions define the cost: finding inputs, coordination, and approvals and verification.
— OpenAI, "The Next Era of Knowledge Work". Previous workplace software lowered the cost of producing artifacts without reducing the attention required to consume them: email made correspondence cheap, then multiplied correspondence; docs made drafting cheap, then multiplied drafts and review cycles. Codex, OpenAI argues, is the redesign.
Up from less than a third in mid-April — and OpenAI calls it the most consequential shift in behavior. Moving from sequential to parallel use lets a single knowledge worker "operate at the scale of a small team," orchestrating work streams instead of executing one task at a time.
Inside Codex you can now highlight the exact part of a document or artifact you want the model to reason over, rather than explaining it in words — extending the interaction model already used for website feedback across all outputs.
Six new function plugins — sales, data analytics, creative production, product design, public equity investing, investment banking — bundle 62 apps and 110 skills, roughly 10 apps and 20 skills per role. It's like packaging the setup of the best user in each function, which makes the plugins a form of product-led education.
Turn any artifact built in Codex into a full shareable website or web app: a revenue-forecast planner instead of a spreadsheet, an event-operations dashboard, a product launch hub. Easier to share than a PDF — just send a URL — and updatable as things change. It's the update NLW is most excited about.
— Simon Smith, Click Health. His read: this puts vibe coding directly in the hands of everyone in an organization — you can build, share, and deploy, and crucially do it in a more secure way, which has been a real issue with internal vibe-coded tools.
Calling this vibe coding is a terminology problem: these are purpose-built, time-boxed tools whose only link to software engineering is that code delivers the output. Just as slide decks, documents, and spreadsheets are core knowledge-work primitives, building small websites and disposable web apps is about to become one too — and Sites makes it radically more accessible.
The company that became exhibit A in the changing tides of agentic AI now has a hard monthly token-spend cap for all employees. Whatever you think of the specific strategy — NLW is saving that for another episode — it's proof that cost management is the other vector of the next wave of enterprise AI.
The headliner is MAI Thinking One: a one-trillion-parameter mixture-of-experts model Microsoft places somewhere in the Sonnet 4.6-to-Opus 4.6 range — trained, per Prime Intellect's Elliot Bakoush, with zero synthetic data or distillation. Sean Wang: Mustafa Suleyman "built a full-fledged neo lab inside Microsoft in two years" that Microsoft controls from chip to model to harness — "absurdly impressive."
— Ethan Mollick, on MAI Thinking One's confusing benchmarks. The skeptics were blunter — leaker I Rule the World called the model "not competitive, particularly not for anything agentic," and Thinking One's scores on Terminal Bench 2.0 and SuiteBench Pro trail Anthropic and OpenAI models from a generation ago.
Don't read these models as something you'd fire up instead of GPT-5.5 or Opus 4.8. Microsoft Frontier Tuning is the point — when Microsoft tuned its models for McKinsey's tasks, MAI delivered the highest win rate, beating GPT-5.5 on quality at 10x lower cost. Given Microsoft's unmatched enterprise distribution, the play is worth taking seriously.
— Satya Nadella, on stage at Microsoft Build. Nadella called it a pretty significant shift — the pitch behind Frontier Tuning is custom, company-specific agents "that only you control," turning enterprises from consumers of the frontier into participants in it.
Two weeks ago a signing ceremony was scheduled and a who's who of tech CEOs invited — then Trump pulled the executive order hours before, saying "I didn't like certain aspects of it," after an eleventh-hour call from David Sacks. This week a substantially identical order was signed in private with zero fanfare.
Safety testing stays voluntary — all major labs have already agreed to submit advanced models — and companies are encouraged to share models 30 days before release, down from the draft's backlash-triggering 90. The NSA leads testing, the Treasury runs a new cybersecurity clearinghouse, and a new disclaimer expressly disavows any mandatory licensing, pre-clearance, or permitting regime.
It reads like the administration eating its vegetables rather than the Big Mac it would prefer: everyone gets something to comment on and claim victory around. As CNAS's David Remler put it, the order "effectively formalizes what has already been happening between the US government and the leading AI companies."
— David Sacks, on which models the order is meant to cover. Sacks acknowledged the FDA-for-AI worry — "bureaucratic mission creep is always a danger, and this should be closely monitored" — but points out the EO expressly forbids a new licensing, pre-clearance, or permitting regime.
— Dean Ball, former White House advisor. Ball calls classifying the details of the voluntary system egregious — if the regulatory thresholds that trigger pre-deployment review are classified, researchers won't know whether what they're training is regulated. His verdict: "not a huge mistake, but a small to medium-sized one."
— Steve Bannon, on the executive order. Bannon says he strongly believes mandatory testing is coming and intends to ramp up the pressure campaign. In a sign of AI's strange political bedfellows, Bernie Sanders wants the same thing: the voluntary order "does almost nothing to protect Americans. Congress must act."
Anthropic is expanding Mythos access into sectors the initial project didn't cover — energy, water, communications, healthcare, computer hardware. The common thread, per Anthropic: a successful attack on each partner's code base could be catastrophic, in most cases affecting more than 100 million people.
Last Thursday, during the Opus 4.8 rollout: a Mythos-level public model in the coming weeks. This week's Glasswing update: general release requires safeguards "that we, and to our knowledge, all other AI developers have yet to develop." The messaging is about as confusing as the government's executive order.
The Information checked in with teams testing Mythos: most are burning through millions of dollars of tokens quickly, and Anthropic is subsidizing use, so firms aren't even paying full cost. Yet many say they're aligning budgets to build their strategy around Mythos once it's broadly available.
The memory maker plans to double manufacturing capacity by the end of the decade — a break from the cyclicality-scarred caution that has kept new plants from being built even as high-bandwidth memory costs more than doubled this year. The chairman says the shortage could last until 2030: "sudden jumps in price can become a problem and actually hurt sustainability."