Brave Search just became the first major search API to accept micropayments from AI agents. Starting this week, agents can pay for individual searches using the x402 protocol, no subscription required.

This matters because current API pricing doesn’t work for agents. Most search APIs charge monthly minimums or per-thousand-query rates. That’s fine if you’re a company running predictable workloads, but agents make sporadic queries. An agent might search once an hour or sit idle for days. Paying $50/month for 30 searches is terrible economics.

x402 fixes this. It’s an HTTP header-based payment protocol that lets agents attach tiny payments to requests. Brave charges $0.02 per search query through x402, paid in cryptocurrency. No account setup, no billing threshold, no monthly fee. The agent constructs a search request, attaches two cents worth of crypto, and gets results back.

For developers building autonomous agents, this changes the cost structure. You pay exactly what you use. If your agent searches five times this month, you pay ten cents total. The friction of API keys and billing portals goes away.

Brave isn’t the first to try micropayments for APIs, but x402 is designed specifically for the agent use case. Traditional micropayment systems require wallets, accounts, or third-party processors. x402 works at the protocol level. The payment is the authentication.

The timing is notable. We’re starting to see real agent deployments, not just demos. Agents that monitor news, compile research, book travel, manage infrastructure. They all need search, but the old “enterprise contract or generous free tier” model breaks down when you have thousands of small agents instead of a few large applications.

Google’s pricing mess

Meanwhile, Google launched Gemini 4 Hydrogen yesterday, the cheapest model in the Gemini 4 family. Simon Willison updated llm-gemini to version 0.37 to support it, and immediately flagged how confusing Google’s pricing has become.

Gemini 4 Carbon and Argon both cost $2 per million input tokens through November 30, then jump to $4 on December 1. Hydrogen launched at similar rates. Output tokens cost $10 per million now, $20 after the deadline. Every model has the same two-tier pricing with the same cutoff date.

This is weird. The whole point of model tiers is differentiated pricing. Carbon is supposed to be the flagship, Hydrogen the budget option. But they cost nearly the same, at least for now. Hydrogen is marginally cheaper, but not enough to matter for most workloads.

Willison’s point is that developers can’t plan around this. If you build something on Hydrogen because it’s cheap, you need to know what cheap means in six weeks. Google’s current answer is “about the same as the expensive one, then twice as much as it is now.”

The December 1 price increase is probably Google testing elasticity. Charge everyone $2 to build adoption, then see who stays at $4. But it makes the model naming almost meaningless. You’re not choosing between cheap and expensive models. You’re choosing between models that will all get expensive soon.

GitHub pushes agent skills format

GitHub published a guide this week on agent skills, their proposed format for packaging reusable agent capabilities. The pitch is “build once, run on any agent.”

Agent skills are structured files that describe what an agent can do. They’re not tools in the traditional sense. Tools are functions with specific implementations. Skills are higher-level descriptions of repeatable work. “Run tests,” “deploy to staging,” “generate API docs.” Things teams do regularly and could teach an agent to do.

GitHub’s format separates the skill definition from the execution environment. The skill file lives in your repo. Any agent that understands the format can read it and attempt the task. The agent brings its own tools, context, and permissions.

This is GitHub’s answer to the portability problem. Right now, if you teach ChatGPT to run your deployment process, that only works in ChatGPT. Teach Claude, start over. GitHub wants you to define the skill once, then have every agent learn from the same file.

It’s early. The format exists, but ecosystem support is limited. GitHub Copilot supports it, obviously. Other agent platforms will need to adopt the same schema. But the idea is sound. Agents should be able to learn from artifacts teams already maintain, not require per-platform configuration.

For developers, this means thinking about skills as documentation. Write down how deploys work, how tests run, how releases get tagged. Make it machine-readable. Then any agent can help with that work.

What shipped

A few other releases worth noting:

AgentRun (GitHub) is a new open-source framework for managed agents that improve themselves over time. It tracks agent performance, identifies failure patterns, and updates agent behavior automatically. Still early, but the self-improvement loop is the interesting part.

IBM’s LLMON made it into Granite 4.1 3B and 8B models. LLMON is a markup language designed for LLMs, meant to replace JSON and XML in agent-to-agent communication. IBM claims it’s more token-efficient and easier for models to generate correctly. The Granite models now support it natively.

Phota Labs launched an AI photo editor that automates the entire editing process. You upload a photo, it applies professional-level adjustments automatically. No sliders, no presets, just AI deciding what good editing looks like. Consumer tool, not developer-focused, but the automation approach is where photo editing is headed.

Claude went down briefly on November 13. Status page shows elevated errors across claude.ai, Claude Code, and the API. Resolved within a few hours. No post-mortem yet.

United Airlines blocked AI agents from booking flights, citing cybersecurity risks. They won’t say what the specific threat model is, but the decision suggests they’re seeing agents attempt bookings and don’t trust the security implications. First major consumer service to explicitly ban agent access.

The Brave Search x402 integration is the most practical development here. Agents need search. Agents need usage-based pricing. This is both. If other APIs follow, we’ll finally have infrastructure that matches how agents actually work.