A glowing turnstile gate stamped with an official seal, one lane opening early, on a navy background

EP 9June 9, 20266 min read

The Government Gets Your AI 30 Days Early + The Token Bill

PodcastPolicyPricing

Matt Wozniak

June 9, 2026 · 6 min read

My take on the executive order that gives the government first look at frontier models, the week your AI bill went usage-based, and the jargon behind it. Companion notes to Human in the Loop Episode 9.

The White House signed an order that lets the federal government test the most powerful models up to 30 days before you can touch them — and the labs said yes. Real national security, or the backdoor the frontier labs were quietly begging for? That's where Oscar and I started, and it only got more interesting from there.

Here's how I called the week.

Signal or Noise

The frontier-model executive order and its 30-day government early-access window

The NSA effectively deciding which models qualify is the part that matters. Early access is influence, and influence flows to whoever holds the on-ramp. Signal.

GitHub Copilot flips to token billing — and bills jump 10x to 50x

This is the one that hit real teams in the wallet. The era of flat-rate AI is ending; your cost is now a function of how much you actually use, and most teams have no idea what that number is. Signal, and act on it now.

Microsoft's MAI-Code-1-Flash and Google's $100 dev tier crash the coding party

Anthropic owned agentic coding; now the giants are undercutting on price. Good for builders. Signal.

ChatGPT crosses 1 billion monthly users

Big number, but the real story is Claude growing ~640% on the back of agentic coding. Noise on top, signal underneath.

Anthropic calls for a global pause — while 80% of its own code is now written by Claude

You don't hand 80% of your codebase to a tool you actually think should be paused. Signal about the gap between what labs say and do.

No Jargon Required

The three terms behind your rising bill:

Tokens — the unit AI is metered in. This is why your bill went usage-based; you're paying per token in and per token out.
Inference vs training — training builds the model once; inference is every time you run it. "Inference-efficient" is the whole game for your cost line.
Evals and benchmarks — why the leaderboard is lying to you, and why your own evals on your own workload are the only score that counts.

The through-line: the money and the influence both moved this week — usage-based to your bill, early-access to the government. Watch or listen below.