RIP Golden Age of Agent Experimentation 2026-2026

May 14, 2026 · Episode Links & Takeaways

Author

AI Daily Brief
May 14, 2026

HEADLINES

Jensen Huang Lands in Beijing

The AI envoy has touched down in Beijing, and Jensen Huang ended up making the flight. Despite Nvidia being a major bargaining chip in US-China relations, Huang was reportedly excluded from the White House's original invite list — only getting the call after his exclusion made headlines and the President personally reached out. The full trade agenda remains uncertain, with Trump likening the summit to Nixon's 1972 meeting with Mao. Commentary is swirling around a potential renegotiation on Taiwan's status, and this week's meeting could have massive ramifications for tech and AI — which we'll cover once a deal is announced.

Cerebras Raises $5.5B in Year's Biggest IPO — After Nearly Getting Acquired

Cerebras raised over $5 billion in the largest IPO so far this year, priced at $185 — well above the advertised range of $150–$160 — implying a market value of $40 billion. Bloomberg reported orders for 20 times more than available shares. As an additional twist, Arm and SoftBank reportedly made a last-minute takeover bid that Cerebras rejected in favor of going public. Hard to second-guess that call now. The stock began trading Thursday morning, so we'll check back in on the Friday show.

Americans Oppose Data Centers More Than Nuclear Plants

A new Gallup poll found that seven in ten Americans oppose data center construction in their local area — opposition so intense it exceeds the all-time high for nuclear opposition at 63%. The top concerns are environmental: excessive resource use, water and electricity consumption, noise, and quality-of-life impacts. Notably, only 12% cited worry about AI replacing jobs, and only 13% cited AI safety. On the flip side, 66% of supporters cited local economic benefit, and 55% specifically mentioned job opportunities — which makes it pretty clear what AI companies should actually be promising. For all the AI advocates complaining about environmental misinformation, there's no amount of being annoyed that's going to change people's minds. As Mads Campbell put it, data centers are fundamentally a marketing problem, and the industry is losing hearts and minds badly.

OpenAI Backs State AI Regulation

OpenAI is now supporting multiple pieces of AI regulation as part of a broader policy repositioning. In a new Axios interview, Chief Global Affairs Officer Chris Lehane acknowledged that AI companies will be crushed by public sentiment if they don't find a way to redistribute AI wealth — drawing a comparison to how Alaska shares oil and gas revenue with citizens. Alongside the interview, OpenAI threw its weight behind regulatory efforts in Illinois, supporting the Kids Online Safety Act and SB315. It seems OpenAI is now on board with state-by-state regulation, as long as it doesn't result in a patchwork of conflicting standards. OpenAI is also floating the idea of an international AI governance body modeled on the IAEA, led by the US and China — though with no OpenAI representative on the Beijing trip, that's unlikely to be on this week's agenda.

Anti-AI Feeling Is More Entrenched Than You Think

A useful reminder of just how entrenched AI opposition is: Schloms on X posted a cropped Monet painting, told people it was AI-generated, and asked them to explain why it was inferior. Hundreds of people obliged with detailed critiques — of a real Monet. The point isn't to dunk on anyone's art knowledge. It's that those of us who think AI can have a positive impact on the future have a lot of work to do.

MAIN STORY

RIP Golden Age of Agent Experimentation

Anthropic's new pricing model for programmatic Claude usage has triggered a developer meltdown — but the real story isn't about Anthropic's relationship with third-party developers. It's about the end of the AI subsidy era. We are crashing headlong into the inevitable market consequence of there being massively more demand for tokens than supply, and this change was completely telegraphed. The question now isn't whether the subsidies end, but how fast.

IS AI’s EXPERIMENTATION ERA OFFICIALLY DEAD?

What Actually Changed
Two pools, one subscription — and a whole new bill for power users.
As of June 15th, Anthropic is splitting paid plan usage into two categories: interactive use (sitting in front of Claude AI, Claude Code, or Claude Cowork) and programmatic use (the Agent SDK, Claude -p, GitHub Actions, third-party tools built on the SDK). Interactive limits are unchanged. For programmatic use, subscribers now get a monthly credit equal to their plan price — $20 for Pro, $100 for Max 5x, $200 for Max 20x — and beyond that, usage is billed at API rates. Anthropic presented this as a clarification and a bonus. Developers who've been running third-party tools on their subscriptions experienced it as a 25–40x cut.

The Developer Crashout
"Any statement from an Anthropic employee is a lie on a timer."
The reaction from developers was furious, particularly from those who had built products specifically using the Agent SDK as Anthropic had instructed. Theo, who built T3 Code, said his users just saw their rate limits cut by 40x despite doing everything right. The biggest specific casualties were Conductor, a popular third-party multi-agent coding harness, and any workflow using OpenClaw. A secondary complaint — arguably louder — was about how the change was communicated. When Anthropic's own Lydia Hallie posted a diagram framing the credit as an "included bonus," the backlash intensified. Gauntlet founder Austen Allred put it plainly: stop dressing up a price hike in 10 layers of PR speak.

Ecosystem Lock-In Is the Strategy
Apple-style control of the end-to-end experience, by design.
Separately from the compute constraints, it's clear that Anthropic has a deliberate bias toward owning the application layer. Just this week they updated Claude for Legal, Claude for Finance, and launched Claude for Small Business. At the same time, they gave Claude Code users a 50% limit increase through July. They are not against third-party tools in principle, but they are not interested in subsidizing competitors. Whether that's the right call is debatable — developers tend to prefer open ecosystems, consumers often prefer the consistency of end-to-end control — but the direction is unmistakable.

Anthropic's Enterprise Business Is Booming
34% of Ramp's customers now pay for Anthropic — up from 9% a year ago.
The pricing power flex makes more sense in light of the enterprise numbers. New data from Ramp found that Anthropic has overtaken OpenAI in business usage for the first time, with 34.4% of Ramp's customers paying for Anthropic vs. 32.2% for OpenAI, with Anthropic's adoption quadrupling year over year. The Information also reported that even mid-size firms are readily signing six and seven-figure Anthropic deals. Almost by definition, the users most affected by this change are net-negative for Anthropic — they cost more than they pay. Cutting back on subsidies to your least profitable cohort, during a compute crunch, was entirely inevitable.

How Big Was the Subsidy, Actually?
One session cost $31 in API tokens — on a $100/month subscription.
The reason it's hard to fully sympathize with some of the complaints: this was extremely well-telegraphed, and the numbers around the subsidy are genuinely staggering. Anthropic's own estimate last month was that the average enterprise developer costs them $13/day — $150–$250/month — just on Claude Code usage. Back in January, Cursor estimated that a $200 subscription allowed $2,000 in compute, with some sources suggesting as high as $5,000. A 10x to 25x subsidy was never a sustainable business model. We are about to find out what AI actually costs.

The Token Maxing Era Is Ending
A brief golden window is closing, and OpenAI won't be far behind.
For a short, remarkable window, you could set an agent to build things overnight as a pure learning exercise with zero commercial intention and basically zero cost. That period is over, interrupted by the laws of physics and supply and demand. The SpaceX compute deal is not a fix — Anthropic has unlocked a few hundred thousand H100s and H200s, but when demand is growing stratospherically and you still have to train the next generation of models, even Colossus-1 is a drop in the bucket. And OpenAI, which has so far avoided making the same move, is likely 6–12 months away from having to do the same. If you're currently using tools Anthropic doesn't love, run to Codex now and soak up that subsidy while it lasts — because it won't.

George from Product Management World (X) The token maxing era is ending
Nityesh (X) OpenAI will follow a similar model within a year