AI in 15 — May 14, 2026

Ninety-nine percent of my usage is non-interactive. That's a Claude Code power user on the Max plan, doing the math on a sudden Anthropic policy change and concluding his bill is about to jump tenfold. The era of all-you-can-eat AI coding subscriptions just hit its first wall.

Welcome to AI in 15 for Thursday, May fourteenth, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Loaded Thursday lineup, Marcus. Anthropic is walling off Claude Code's headless mode behind a new metered credit pool — and developers are furious. Closing arguments land today in Musk versus OpenAI after Sam Altman's bruising testimony. Anthropic launched Claude for Small Business and a ten-city US tour. Nvidia tied itself to David Silver's billion-dollar reinforcement-learning lab. Meta made its new AI chatbot account on Threads unblockable. Medicare quietly built a ten-year payment model engineered for AI agents. And BeyondTrust dropped a serious security disclosure on OpenAI Codex.

Anthropic meters headless Claude Code — and the developer reaction is intense.

Nvidia bets on superlearners.

And Meta ships a chatbot you cannot block.

Lead story, Marcus. Anthropic dropped a Claude Code policy change yesterday that has the developer community in revolt. Walk me through it.

Real retrenchment, Kate. Starting June fifteenth, any programmatic use of Claude Code — the headless dash-p flag, the Agent SDK, and third-party agent wrappers like OpenClaw — will no longer draw from the general subscription pool that interactive users tap. Instead, those uses get a separate Agent SDK credit allocation, sized at twenty to two hundred dollars of API-equivalent per month depending on the plan tier, billed at full API rates and non-rollover. Once you exhaust the credit, programmatic calls stop unless you've enabled extra usage at pay-as-you-go API prices.

And the official framing.

Anthropic is calling this a reinstatement of third-party agent access with clearer rules, Kate. OAuth credentials are for interactive personal use. Anything built on top — agent platforms operating on your behalf — should run through API keys. But the developer math is brutal. Power users on the two-hundred-dollar Max plan report that ninety-nine percent of their usage is non-interactive — CI workflows, voice-driven coding over Tailscale, batch jobs. One developer on Max five-x calculated his current forty-percent usage would translate to roughly a thousand dollars a month under the new scheme. A ten-x price increase overnight. Two hundred forty-three comments on Hacker News, and threads of users already switching to OpenAI's Codex.

Why does this matter beyond developer grumbling.

Because it's the first real crack in the all-you-can-eat AI coding subscription model, Kate. For twelve months Anthropic told developers the Claude Code subscription was the most generous deal in AI coding. It tells you something about unit economics that even after the SpaceX Colossus compute deal doubled headline limits, agentic and headless workloads are evidently the heaviest part of the load — and Anthropic cannot keep subsidizing them out of a flat fee. It also draws a sharp line in the emerging coding agent market. Third-party wrappers now have to plug into API billing rather than ride someone's Max plan. Expect a similar move from OpenAI on Codex within the year. The libertarian read, Kate, is the market is finally finding the real price of compute. Subscriptions priced like all-you-can-eat buffets were never sustainable.

Quick hits. Marcus, closing arguments in Musk versus OpenAI are today. As we covered yesterday, Altman took the stand. Anything new.

The follow-on testimony filled in the picture, Kate. Altman told the jury Musk had tried to kill OpenAI twice during the original split, and that when Musk walked, the nonprofit had been, quote, left for dead. Earlier in the trial Musk himself acknowledged on the stand that xAI distills OpenAI's models — which is awkward for a self-described independent challenger. Microsoft witnesses revealed the company quietly feared becoming too dependent on OpenAI. Closing arguments today, then a nine-person jury deliberates. The verdict could affect OpenAI's cap table, Altman's CEO seat, and Microsoft's thirteen-billion-dollar investment all at once. The judge's ruling lands next week.

Anthropic story, Marcus. Claude for Small Business launched yesterday.

Clean strategic pivot, Kate. Anthropic unveiled a packaged product targeting Main Street — connectors into QuickBooks, PayPal, HubSpot, Canva, DocuSign, Google Workspace, and Microsoft 365, plus fifteen pre-built agentic workflows. Payroll planning, monthly financial close, invoice chasing, margin analysis, marketing campaigns, onboarding, tax prep. Every action requires explicit user approval before it runs — Anthropic clearly learned from the early agent-runs-amok stories. The product sits inside Claude Cowork, their task automation platform. And starting today they're kicking off a ten-city tour — Chicago, Tulsa, Dallas, New Jersey, Baton Rouge, Birmingham, Salt Lake City, Baltimore, San Jose, Indianapolis — with free half-day AI Fluency workshops for a hundred local owners per city.

Daniela Amodei pitched it directly.

Quote, Claude helps take the late-night work off their plates, Kate. The math behind it is real. Small businesses are forty-four percent of US GDP and roughly half of private-sector employment, but AI adoption in that segment has trailed enterprise badly. Anthropic is betting the wedge is workflow templates, not a smarter chatbot. The thirty-person landscaper doesn't want a model. He wants his books reconciled. Strategically, this opens a second front that doesn't compete head-on with OpenAI's enterprise sales motion or Microsoft's Copilot bundling. With Anthropic reportedly raising at a nine-hundred-fifty-billion-dollar valuation, courting the long tail of US businesses is how you justify those numbers without depending solely on the Fortune 500. PayPal partnership and CDFI outreach round out the package.

Nvidia story, Marcus. They partnered with Ineffable Intelligence yesterday.

Significant bet, Kate. Nvidia announced a multi-year technical partnership with the London-based lab founded by AlphaGo and AlphaZero architect David Silver. The two companies will jointly build reinforcement-learning infrastructure designed for systems that act, observe, score, and update continuously in tight loops — a workload pattern very different from the static-dataset pretraining that produced today's LLMs. Engineering teams are co-locating. Work starts on Grace Blackwell and is intended as a lead workload for the upcoming Vera Rubin generation. Jensen called it the next frontier — quote, superlearners, systems that learn continuously from experience. Silver was blunter — researchers have largely solved the easier problem of AI, how to build systems that know all the things humans already know.

So Nvidia is hedging beyond the LLM paradigm.

Pretty explicitly, Kate. Reinforcement learning at scale has been treated as a sidecar to pretraining for the last three years. If Silver and Jensen are right that the next jump comes from agents learning in simulators rather than reading the internet, the hardware bottleneck shifts from raw FLOPs to interconnect bandwidth and serving latency — which is precisely what Blackwell and Rubin are optimized for. Ineffable raised a record one-point-one-billion-dollar seed in April from Sequoia and Lightspeed, with Nvidia, DST, Index, Google, and the UK Sovereign AI Fund participating. The geopolitical layer is worth noting. A credible non-US frontier lab now exists outside Big Tech proper, headquartered in London, on Nvidia silicon. If Ineffable shows benchmark results before year-end, the scaling-laws-are-dead debate gets a hard new data point.

Meta story, Marcus. They made their new Threads AI account unblockable.

Genuinely tone-deaf product decision, Kate. Meta is testing a Grok-style AI chatbot on Threads that posts publicly, replies to threads, and can be summoned with an at-mention. They stripped the standard block button from the at-meta-dot-AI account. Users found the gap immediately — reporting for spam doesn't surface a block option, and existing block attempts have no effect. Meta's response — you can mute, hit not-interested, or hide replies. But you cannot block. The chatbot beta is live in five countries — Malaysia, Saudi Arabia, Mexico, Argentina, Singapore — but the account itself is visible everywhere. Quote, users cannot block Meta AI became the top trending topic on Threads with over a million posts.

Why does this matter beyond the outrage.

First time a major platform has hard-coded an AI presence users cannot refuse, Kate. That's a real design precedent. And it's a tell on Meta's strategy. With a hundred fifteen to a hundred thirty-five billion in 2026 AI capex, they need engagement numbers that justify the buildout — and an unblockable bot is the most aggressive possible answer to where's-the-demand. The libertarian objection writes itself. Meta is using monopoly platform power to override consumer choice on an unproven product. Expect copycat moves from X with Grok and from LinkedIn, and probably an EU regulatory inquiry once this expands beyond the five test markets. Comparisons to Bluesky's AI assistant — which became the second-most-blocked account on that platform — are everywhere.

Healthcare story, Marcus. CMS launched something quietly that the tech press is missing entirely.

This is the sleeper story of the week, Kate. CMS will launch ACCESS — Advancing Chronic Care with Effective Scalable Solutions — on July fifth, a ten-year payment model that flips Medicare's reimbursement logic. Instead of paying for clinician minutes, ACCESS pays organizations a predictable per-patient amount for managing chronic conditions like diabetes, hypertension, CKD, obesity, depression, and anxiety. Full reimbursement is contingent on hitting measurable outcomes — lower blood pressure, reduced pain. One hundred fifty organizations were selected, including AI-first care companies like Pair Team, which uses a twenty-four-seven voice agent called Flora for intake, check-ins, and housing and medication coordination.

And the structural change.

Crucial, Kate. Under fee-for-service Medicare, there is literally no billing code for, quote, an AI agent that calls a patient between visits. ACCESS creates that mechanism for the first time by paying for outcomes rather than activities. The reimbursement rates came in lower than expected, which means only operations with substantial automation can make the math work. Exactly the kind of natural selection for AI-native care models that CMS clearly wants. Execution risk is real — the 2023 CMS Innovation Center added five-point-four billion in costs rather than producing savings. But roughly a trillion dollars of US healthcare spending flows through Medicare. If even a fraction reroutes through outcome-based contracts, the moat for traditional EHR vendors and fee-for-service telehealth collapses. And the political angle is clean. Government creating a price signal rather than dictating an AI mandate. That's the libertarian-favored mechanism right there.

Security story, Marcus. BeyondTrust disclosed a serious Codex vulnerability yesterday.

First major published exploit chain against a frontier-lab coding agent, Kate. BeyondTrust's Phantom Labs found a command-injection flaw in OpenAI Codex that let an attacker steal GitHub OAuth tokens by crafting branch names containing payloads, including obfuscated Unicode characters for stealth. The Codex agent executed those branch names in its container, exposing repository, workflow, and private-code tokens. The disclosure also detailed potential lateral movement across organizations sharing Codex environments, and demonstrated automated exploitation at scale. Reported to OpenAI in late December. Patched in current Codex versions. The researchers' framing — quote, AI coding agents are not just productivity tools — they're execution environments needing the same security rigor as any other application boundary.

And there's a separate Composer disclosure too.

Patched yesterday, Kate — leaks GitHub Actions tokens into CI logs. Different vector, same theme. Coding agents rocketed from a niche dev tool to running real production code in months, but the security model is borrowed from chat UIs, not from CI/CD. This exploit will be cited for the next two years by anyone arguing that agentic AI needs sandboxing built in, not bolted on. The practical takeaway for listeners — scoped short-lived tokens aren't enough. Agents must be assumed compromised and isolated accordingly.

Big picture, Marcus.

Three threads tie today together, Kate. First — the honeymoon pricing is ending. Claude Code's metered credits, the Codex security tax, and the Composer-style supply-chain bugs all point the same direction. The first wave of, quote, use AI agents however you want is being replaced by metered, audited, billed-per-call reality. Compute economics and security debt are both catching up. Second — AI commercialization is moving downmarket and into infrastructure. Claude for Small Business hits Main Street. Medicare ACCESS embeds AI agents into healthcare reimbursement. Nvidia and Ineffable bet on the substrate beneath the next generation. Lots of plumbing being laid this week, less benchmark theater. Third — the trust crisis is becoming literal. Altman in the witness box answering, are you completely trustworthy. Meta refusing to let users block its bot. Developers reporting their bills are about to ten-x. The AI industry's reputation, with users, courts, and its own workforce, is now a first-class business problem, not a PR one. The pro-Western libertarian read, Kate, is that prices finding their level, courts holding labs to their charters, and outcomes-based payment models like ACCESS are all healthy market signals. The cautionary thread is Meta's unblockable bot. That's the one move on today's list that depends on monopoly power rather than competition, and it's the one that should get regulatory attention.

That's your AI in 15 for today. See you tomorrow.