It feels a lot like companies are starting to turn off the inference spigot.
The last hype cycle around AI was around tokenmaxxing. Learning about the tools by using as many tokens as possible. Let the agents run, don't worry about efficiency. More output means you'll eventually get better output.
But the bill is starting to come due. GitHub is moving Copilot to usage-based billing on June 1. Same monthly price, but now your usage is measured in tokens and credits instead of requests. Anthropic may have been experimenting with pulling Claude Code off the $20 Pro plan (to much outrage). Hell, we even blew through our OpenAI monthly credit usage for the first time on our Enterprise plan with 10 days to spare. Usage costs are on the rise.
Armin Ronacher and Ben Vinegar are calling it "the beginning of the end of subsidies" on their podcast this week, and that puts to words what I've been feeling creep in. The seat-based pricing models that worked six months ago can't support the way agents actually use compute, so everyone's switching to usage-based billing. And usage, it turns out, costs a lot when you're running things on overdrive.
If we're moving on from tokenmaxxing, what does that mean for the next phase? My theory is we'll see more focus on building efficient systems around AI. We're entering the agentic engineering phase. We're starting to understand what agents can do and now we need to figure out how to make them efficient at doing it.
It'll be interesting to see how Agents as a Service vs open-source infrastructure take shape. Smaller teams and those looking to be efficient aren't going to be looking to spend an extra $250,000 per engineer in token costs.
It's time to start building systems.
📺 Watch
How I created OpenClaw, the breakthrough AI agent
TED / Peter Steinberger
Peter Steinberger took the TED stage to tell the story of how OpenClaw went from a WhatsApp bot he hacked together to one of the fastest-growing open-source projects in history. I'm very interested in how OpenClaw could be utilized with local models to keep costs in check as the cost of compute rises.
🎧 Watch
State of Agentic Coding #6 with Armin and Ben
YouTube
Armin Ronacher and Ben Vinegar have a sprawling conversation about agentic engineering. They break down their takes on the rising costs for AI tools and how provider and user incentives may not perfectly align. At AI Engineer Europe, an OpenAI engineer told the crowd to just let the agent run and not worry about token spend. The very next talk was Mario Zechner explaining why efficiency is the whole point. They also dig into how AI security harnesses need to evolve as agents find real vulnerabilities in production code. It's the most grounded take I've heard on what agentic work actually costs right now.
🔧 Tool
Pi Coding Agent
pi.dev
Pi is an open-source project for building your own agentic harness. It's minimal by design, MIT-licensed, and works with 15+ model providers. As personal agents become more common, Pi could prove to be the tool to reach for when you need specific, focused agents without the overhead of a full platform. Armin Ronacher's company Earendil recently acquired Pi, which tracks with the efficiency-first mindset they're preaching on the podcast.
📖 Read
WordPress is my Open Claw now
Ronnie Burt
I've had the idea in the back of my head to use WordPress as a knowledge base for agents, and Ronnie Burt at Automattic has an interesting proof of concept with a similar approach. He built HelloDolly.fun on top of WordPress Multisite as an accessible agent platform. I'm very interested in seeing how WordPress is able to fit itself into the agentic landscape. It's never been the sexiest platform, but it still provides a lot of built-in functionality in an open-source package.
SpaceX struck a deal with Cursor to build a "coding and knowledge work AI," with an option to acquire the company for $60 billion. It makes sense. Cursor needs compute resources which SpaceX has, and xAI needs to catch up in the coding game, which Cursor has. Cursor has been putting out some interesting things lately. I’m keeping an eye on them going forward.
Thanks for reading, Jason
