🤖 AI News Roundup: The AI Race Lead Shifts

The centre of gravity in AI moved this week, although not at the model layer where headlines usually sit. Anthropic took the lead in paid business adoption for the first time on Ramp's index, while OpenAI spun out a $14 billion deployment subsidiary and Anthropic restructured its agent pricing into a separate credit pool. Bloomberg confirmed that AI-exposed occupations have now posted two consecutive years of decline while the rest of the labour market grew, and a Nobel-winning economist named the markers that decide whether AI actually changes work. Below, eight stories on where this is going and what to watch.

OpenAI Spins Out a $14B Consulting Arm and Buys Tomoro to Staff It

On 11 May 2026 OpenAI launched the OpenAI Deployment Company, internally called DeployCo, structured as a majority-owned subsidiary with around $4 billion of fresh capital at a $10 billion pre-money valuation, putting the new vehicle at roughly $14 billion post-money. TPG leads the investor group alongside Advent, Bain Capital, Brookfield, Goldman Sachs, SoftBank, and the consulting firms Bain & Company, Capgemini, and McKinsey, all of whom are on the cap table. Alongside the launch OpenAI agreed to acquire Tomoro, a London-headquartered AI engineering firm founded in 2023 in alliance with OpenAI, bringing in roughly 150 forward-deployed engineers to staff DeployCo's client work. Beneath the announcement language about helping organisations "build around intelligence", OpenAI has stood up a professional services and consulting arm to sit alongside its model business, positioning to capture implementation margin that has, until now, mostly flowed to systems integrators. For leaders, the build-internally, buy-from-a-consultancy, or let-the-vendor-run-it choice now has a fourth option attached to it, where the vendor's own deployment arm runs the integration, and the trade-offs between speed, lock-in, internal capability building, and long-term cost need working through rather than waving away. The vendor-dependency concerns that come with consolidating around a single model provider become harder to ignore once that provider is also running your projects.

Anthropic Splits Agent Usage from Subscription Limits with a Monthly Credit Pool

Anthropic announced on 14 May that from 15 June, Agent SDK usage and the claude -p non-interactive command will no longer draw from Claude subscription rate limits. Subscription holders instead receive a separate monthly programmatic credit pool: $20 for Pro, $100 for Max 5x, $200 for Max 20x, $20 per seat on Team Standard, and $100 per seat on Team Premium, with credits that do not roll over and metering at API list rates once exhausted. Interactive Claude Code, Claude Cowork, and the chat interface stay on existing subscription limits. The change closes a loop Anthropic opened in April when it blocked third-party agents from running on subscription rates. Theo Browne pointed out on X that describing the new allocation as "free credit" obscures the removal of what had effectively been a 25x subsidy on programmatic usage, and that downstream products built on Claude subscriptions will now have to throttle agentic features to stay inside the new caps. The wider point worth holding is the shift in pricing model, where per-seat licensing assumes a human at the keyboard while per-agent-task pricing assumes the work is being done by software the licence holder dispatched. Leaders who signed flat-rate Claude subscriptions on the assumption that agentic automation was included should expect to renegotiate that across vendors over the coming year.

The EU AI Act's Article 50 Transparency Rules Take Effect on 2 August 2026

Article 50 of the EU AI Act sets out the transparency obligations for anyone placing an AI system on the EU market or deploying one inside the bloc, with the full set of rules taking effect on 2 August 2026. The duties split into two camps. Providers, meaning the organisations that build and supply AI systems, must design those systems so that users are told they are interacting with AI and so synthetic outputs carry a machine-readable mark identifying them as artificially generated. Deployers, meaning the organisations putting those systems to use, must inform individuals when they are subject to emotion recognition or biometric categorisation, and must disclose when content has been generated or manipulated by AI in matters of public interest, including deepfakes. In practice a chatbot must declare itself a chatbot at the start of the conversation, a generative image tool must embed a watermark, and a platform publishing AI-generated text on a current-affairs matter has to label that text unless a human editor has taken substantive responsibility for it. The standard is a step beyond familiar notice-and-consent practice, because the obligation attaches to the act of producing or deploying the system rather than to whether the user agreed to anything. For leaders shipping anything into the EU, watermarking pipelines, chatbot disclosure copy, and review processes for AI-generated public content all take time to put in place. Article 50 is notable because it binds at the deployment layer rather than the model layer, where most of the real-world exposure actually sits.

AI Cyberattack Capability Is Doubling Every 4.7 Months, and the First AI-Discovered Zero-Day Lands the Same Week

Two reports published in the same week confirm the same trajectory from independent angles. The UK AI Security Institute, in its May 2026 update on autonomous cyber capability, reports that the length of cyber tasks frontier models can complete unassisted is now doubling roughly every 4.7 months as of February 2026, accelerating from an eight-month doubling rate estimated only the previous November. Claude Mythos Preview and OpenAI's GPT-5.5 both exceeded the existing test suite, with Mythos Preview the first to solve both of AISI's cyber-range exercises and both models achieving near-total reliability on tasks estimated to take an experienced human eight hours or more. In the same window, Google's Threat Intelligence Group disclosed what it describes as the first case of a threat actor deploying a zero-day exploit developed with the help of an AI model. The target was a 2FA bypass in a widely-used open-source web administration tool, and GTIG attributes AI involvement to telltale code characteristics including textbook Pythonic structure, abundant educational docstrings, and a hallucinated CVSS score in the exploit script itself. AISI is measuring the slope in controlled evaluations while GTIG is observing the same slope appear in the wild, and both reports converge on a framing of compressed defender economics. For small business owners the implication is sharp, because the standard objection that an organisation is too small or too unremarkable to be a target loses what little weight it had once vulnerability discovery is automated, at which point targeting follows the path of least resistance.

Anthropic Overtakes OpenAI in Paid Business Adoption for the First Time

For the first time on Ramp's index, Anthropic is sitting ahead of OpenAI on paid business adoption, reaching 34.4% of the businesses Ramp tracks in April 2026 after a 3.8-point monthly rise, while OpenAI fell 2.9 points to 32.3%. Over the past twelve months Anthropic's share has climbed from around 9% to 34.4%, while OpenAI's paid share has barely moved. The index draws on corporate card and invoice payment data from roughly 50,000 US businesses on the Ramp platform, which gives it a near real-time read on what companies are actually paying for rather than what they say they intend to buy. Ramp's economist Ara Kharazian frames the swing as the visible result of a longer strategy, describing Anthropic's approach as starting with a technical customer base and broadening through developer-facing tools. Claude Code looks like the anchor of that broadening, because strong adoption inside engineering teams creates a route into the wider organisation once procurement, finance, and legal start sharing the same vendor relationship. For anyone choosing how to deploy these tools day-to-day, the practical reading is that the market has stopped being a one-name conversation, and workers may find their organisation now standardising on Claude alongside or instead of ChatGPT, particularly where engineering teams led the procurement. Leaders setting AI strategy should treat this as confirmation that vendor concentration is a meaningful operational risk and that workflows worth keeping should remain portable across providers.

AI-Exposed Occupations Post a Second Year of Decline as Customer Service Drops Nearly 5%

Customer service representatives lost 130,180 positions over the year to May 2025, a 4.8% contraction in a category that has been on most AI-exposure shortlists since 2022. Bloomberg, working from US Bureau of Labor Statistics data, reports that the 18 occupations the BLS has identified as exposed to artificial intelligence, covering roughly 10 million jobs, saw employment fall 0.2% between May 2024 and May 2025 while overall US employment grew 0.8% over the same window. Stripping out the fast-growing medical secretaries and assistants category, the remaining 17 AI-exposed occupations contracted by 1.6%, the second consecutive annual decline of that magnitude. Since May 2022, the steepest cumulative drops are credit authorisers and clerks (down 26.2%), broadcast announcers (down 20.8%), and sales engineers (down 13.2%), all categories where the work product is now within range of what current generative and analytical models can produce. Two consecutive years of contraction in the same 17 categories, while the broader labour market continues to expand, is hard to read as anything other than structural displacement linked to the technologies these occupations were identified as exposed to. For workers in customer service, secretarial, and sales roles, the displacement many have been hearing about for two or three years is now showing up in employment data rather than in opinion pieces, which means retraining, lateral moves, and skills broadening have stopped being optional planning exercises to revisit later in the decade. The risk of training your own replacement is no longer hypothetical, and workforce planning conversations about which roles will exist in their current form in 2027 and 2028 cannot reasonably be deferred any longer.

Acemoglu Names the Three Markers That Decide Whether AI Actually Changes Work

Daron Acemoglu, who shared the 2024 Nobel in economics for work on the institutional and technological roots of inequality, has set out three markers worth watching to judge whether AI is actually reshaping employment or simply being talked about as if it were. The first is whether AI agents can handle the orchestration between tasks that humans manage without thinking. Acemoglu uses the example of an x-ray technician moving across roughly thirty interconnected tasks in a normal shift, from patient histories to organising mammogram archives, where agents that handle one task well still fall down when the work involves stitching many tasks together, and that orchestration gap is what determines whether a job is partially augmented or actually displaced. The second is that the major labs are now hiring economists in significant numbers, with OpenAI bringing in Ronnie Chatterji as chief economist, Anthropic convening an economic advisory council, and Google DeepMind appointing Alex Imas as director of AGI economics. Acemoglu reads this as a tell, since the labs have noticed that the labour-effect debate shapes their licence to operate and they have an incentive to fund research that points in a favourable direction. The third is whether AI produces the sort of practical, user-friendly applications that earlier general-purpose technologies eventually produced, because PowerPoint and Word landed when anyone could pick them up and get something useful out in minutes, whereas most current AI tools still demand a learning curve before they pay off. For workers, this is the variable set that decides whether your job actually changes, because the question is not whether a lab demos something impressive but whether the agent can move across the messy task graph your career already runs on.

Koomen Argues the Corporate Hierarchy Is a Routing Hack That AI Is Now Collapsing

The argument running through Pete Koomen's piece is that corporate hierarchies are an information-routing system inherited from Roman military structure, built because shifting decisions between people was historically more expensive than doing the underlying work, and that AI is now collapsing the routing layer faster than the labour market has language for. Koomen describes organisations as graphs of nodes (the actual work) and edges (the handoffs between functions), where the edges were always more expensive than the nodes, so an unusually large share of corporate headcount sits in roles that exist to relay information, precompute decisions, and translate between layers rather than to produce a work output directly. Koomen is worth listening to on this because he is a practitioner rather than a commentator, having cofounded Optimizely and now working as a partner at Y Combinator, which puts him close to the operators who are actually restructuring around AI inside both startups and larger firms. The mechanism he sets out is that AI agents simultaneously automate the routine work that used to fill an individual contributor's day and the coordination work that used to fill a middle manager's day, which closes both ends of the role from the same direction. He points to Block's February 2026 reorganisation, where the company cut from roughly 10,000 to under 6,000 employees and rebuilt the structure around three explicit roles, with traditional management layers removed entirely and the stock rising 25% on the announcement. He cites Meta's Reality Labs running AI-native pods of up to fifty individual contributors per manager, with target adoption at 75% AI-written code across 65% of engineers, and the result on the org chart is that the role of project manager, QA analyst, data analyst, and similar coordination positions falls inside the squeeze while the technical contributors and the small layer that sets direction sit outside it. For workers in pure coordination roles, the question worth asking is what proportion of a normal week is spent producing a direct output a stakeholder consumes versus moving information between people who could now move it themselves with an agent in the loop. The middle-management compression Koomen describes is the mechanism by which the Bloomberg data and Acemoglu's orchestration marker actually arrive inside firms, and for leaders restructuring around AI the design assumption that needs to change is that coordination headcount scales linearly with throughput, because the Block and Meta examples suggest the new ratio is closer to a thin coordination layer plus deeper individual-contributor capacity.