Anthropic's Run-Rate Math and What the $26M Question Actually Means

GitHub Copilot just blew up its pricing model, and the dev community noticed immediately.

Microsoft flipped Copilot from a flat monthly fee to token-based billing this week. The reaction was loud and mostly unprintable. Developers who built workflows around predictable monthly costs are now watching token meters tick up on tasks they assumed were covered. The complaints filling GitHub Discussions and tech Twitter have a specific flavor: this isn't frustration at a price increase, it's frustration at losing the floor.

That story matters beyond one product. Anthropic is running at a pace that has Reuters columnists questioning how the math holds together. SoftBank is committing €75 billion to French data centers. The infrastructure money is moving fast, and the billing models are still being invented in real time.

While the big players fight over who owns the meter, this issue is about what you can build and ship. Eight repos, two skills, three stack picks, and Friday's ClauseGuard kit starts teasing today. Let's get into it.

---

1. GitHub Copilot goes token-based and devs revolt. Microsoft switched Copilot to consumption billing, ending the flat-fee era for many plans. Developers are calling it a pricing trap for heavy users. [TechCrunch]

2. Anthropic's revenue math is messier than reported. Simon Willison breaks down how Anthropic calculates "run-rate revenue", a 28-day consumption figure multiplied by 13, plus monthly subscriptions times 12. Karen Kwok at Reuters Breakingviews questions whether that framing holds up. [Simon Willison]

3. SoftBank commits €75B to French data centers. The target: 5 gigawatts of new capacity. That's not a pilot program. AI infrastructure spending is no longer a Silicon Valley story. [TechCrunch]

4. Anthropic publishes how it sandboxes Claude across products. Willison flags a rare piece of thorough documentation on Claude's containment architecture, what's blocked, what's allowed, and why. Worth reading if you're building anything that ships Claude to end users. [Simon Willison]

5. MiniMax M3 hits Vercel AI Gateway with a 1M-token context window. M3 is MiniMax's first multimodal model, built around sparse attention. Now routable through Vercel's unified gateway alongside the major providers. [Vercel Blog]

---

Repos

[Repo] LangGraph, Graph-based agent orchestration from LangChain. Lets you build stateful multi-step agent loops where each node is a function and edges define transition logic. If you're running agents that need to retry, branch, or resume mid-task, this is the framework that doesn't make you build that infrastructure from scratch.

[Repo] Aider, AI pair programmer that runs in your terminal and knows your git history. Makes multi-file edits, understands your project structure, and has an architect mode that plans changes before touching code. The practical replacement for context-copying into a chat window.

[Repo] Vercel Sandbox with Docker, Vercel's sandbox now runs full Docker containers, letting agents install system packages, build containers, and modify files without touching the host. Agentic workflows that need an isolated execution environment just got a managed path to production.

[Repo] OpenAI Rosalind Biodefense, OpenAI's new domain-specific model access tier for biodefense and public health. Vetted developers and U.S. government partners can now build on GPT-Rosalind for pandemic preparedness applications. Signals where specialized vertical AI access is heading.

[Repo] bunnyos/base-agent, The first open-source Base agent implementation. Early-stage but notable for anyone watching the agent framework space, Base's primitives are getting community wrappers fast. Clone it before the star count catches up.

Skills

[Skill] Claude containment pattern, Anthropic's documented sandboxing approach for multi-product Claude deployments. Study the specific allow/deny patterns if you're building anything that routes Claude to external users. The documentation gap on this has been a real problem; this fills it.

[Skill] AI model evaluation playbook, OpenAI's shared framework for third-party model evals. Covers how to assess capability, safeguards, and validity for frontier systems. Useful if you're building internal benchmarks or vetting models before production use.

---

[MCP] LangGraph MCP Integration, LangGraph exposes its graph runtime as an orchestration target for MCP-aware agents. Wire Claude to LangGraph as the execution layer and you get stateful, resumable agent loops without managing state yourself. The config trick: define your graph nodes as typed functions, Claude's tool-use maps directly to node transitions with zero glue code.

[Tool] Vercel AI Gateway, A unified API gateway that routes across OpenAI, Anthropic, MiniMax, and other providers. MiniMax M3's 1M-token context window is now available here alongside the major providers. Use it when you want model-switching without rewriting your API calls.

[Tool] Vercel Sandbox, Managed code execution environment that now runs Docker containers inside isolated sandboxes. An agent can build containers, install packages, and run arbitrary code without touching your host. The non-obvious use: give Claude a Vercel Sandbox as its execution target and you have a safe, repeatable agentic build environment in under an hour of setup.

---

Pattern: Containment-layer prompting for Claude in multi-user products

Anthropic published a rare piece of documentation this week on how they contain Claude across products. The core pattern is simple and underused: define your allow-list and deny-list at the system prompt level, not the application level.

The command pattern looks like this:

You are operating inside [ProductName]. ALLOWED: [specific capabilities list] NEVER: [specific off-limits actions, even if asked] If a user request falls outside ALLOWED, respond with: [specific fallback phrase]

Most builders bury restrictions in vague language ("be helpful and safe"). Anthropic's approach names exact allowed actions and exact fallback behavior. The result is a model that fails predictably instead of failing creatively. Apply this before you ship any Claude-powered tool to end users, the documentation is worth reading in full.

---

GitHub Copilot's billing switch tells you something worth naming: flat-rate pricing was a land-grab, not a business model. Token billing is where every AI product eventually ends up, because it's the only structure that matches cost to usage at scale. The question isn't whether your tools go token-based, it's whether your workflows are built to stay efficient when they do.

Build like the meter is already running.

---

This week's kit: ClauseGuard

There is a clause in most standard client contracts that lets your client keep your work without paying the final invoice. It is usually on page 3 or 4, written in paragraph three of the "Intellectual Property" section, and most freelancers sign it without reading it. This Friday's kit, ClauseGuard, is a self-serve contract scanner built entirely on Claude, paste or upload any MSA, SOW, or NDA and get a plain-English risk report in under 60 seconds. This week we're covering the what: what this product does, why Claude is the right engine for it, and why a $400/hr problem is now a $19/mo product. Full build specs, prompts, and the pricing model drop Friday, grab the upgrade to be ready.

---

That's the Monday issue. The Copilot billing story is worth watching, it's the first major sign that flat-rate AI tooling was a temporary subsidy, not a permanent price point. Build accordingly.

See you tomorrow.