In April we wrote that flat-rate AI died for builders — three pricing moves in seventeen days, and a prediction at the end: OpenAI and Google follow with similar repricing on their agent products by summer.

It didn't take until summer. Between June 1 and June 18, four vendors are resetting how coding agents get billed. April was the vendors refusing to subsidize other people's agents on flat-rate plans. June is the vendors putting meters on their own.

June 1: GitHub Copilot moves to AI Credits

The April sign-up pause made sense the moment this shipped. Copilot has moved to usage-based billing: every plan now comes with a monthly allotment of "AI Credits" that roughly matches what you pay — Pro ($10/month) gets $10 of credits, Pro+ ($39) gets $39, Business gets $19 per user. Burn through them and you either top up or stop.

The plan prices didn't change. What changed is what they buy. Early reports describe bills jumping as much as 25× for heavy agent users, with a single long agentic session consuming up to $40 in credits. Tab-completion users will barely notice. Agent Mode users are the entire point of the change.

There's a cushion: a promotional credit bonus through August gives Business seats $30/user/month and Enterprise $70/user/month instead of their face value. Note what a promotional cushion is — a ramp, not a reprieve.

June 1: Codex's promo allowance quietly expires

Same day, lower headline: OpenAI's promotional 2× usage multiplier on Codex Pro plans expired. Message limits dropped from the promoted ~160–800 per five-hour window back to the standard 80–400. No price change, no announcement blog post — just half the allowance you'd gotten used to.

This is the pattern worth internalizing: promotional capacity is marketing spend. When the land-grab phase ends, it gets reclaimed, and your workflows are suddenly sized for an allowance that no longer exists.

June 15: Claude Code gets its own credit pool

The one circled on our calendar. Starting June 15, programmatic Claude usage — Claude Code sessions included — moves off the shared subscription pool onto a separate monthly credit pool billed at full API rates: Pro gets $20 of credit, Max 5x gets $100, Max 20x gets $200. At Opus 4.7's $5/$25 per million tokens, a credit pool is not a big bucket for an agent that loops.

This completes the arc that started April 4. First Anthropic cut third-party frameworks like OpenClaw off the flat-rate plans. Now the first-party agent gets its own meter too. The subscription increasingly buys you the chat surface; the agent surface is metered compute, whoever wrote the agent.

From the workshop floor: this is the change that touches us most directly, and the discipline we wrote about in April — prompt caching as a default, Opus only where it earns its rate, budgets per workflow — goes from "good hygiene" to "the difference between a $100 month and a blown pool by the 9th." If you're on a Max plan and run agents on cron, do the math before the 15th, not after.

June 18: Gemini CLI goes enterprise-only

The bluntest move of the four: Google is shutting down Gemini CLI for free and individual users entirely. Enterprise-licensed customers keep access; everyone else is pointed at Antigravity CLI, which runs on a compute-quota system that refreshes every five hours until you hit a weekly ceiling.

Where GitHub and Anthropic repriced, Google segmented. The free individual tier of an agentic CLI turned out to be a cost center with no obvious path to revenue, so it's gone. If your tooling depended on it, you have until the 18th.

Credits are the new pricing primitive

Look at the shape of all four moves. Nobody raised sticker prices. Instead, everybody introduced or tightened an intermediate currency — credits, pools, quotas — sitting between your subscription and the compute you consume.

  • GitHub: AI Credits, face value pegged to your plan price
  • Anthropic: monthly credit pools at API rates for programmatic use
  • OpenAI: message windows, promo multipliers that expire
  • Google: compute quotas on five-hour refresh with weekly ceilings

Credits do two jobs for the vendor. They cap the downside of agentic power users without headline price increases. And they make comparison shopping nearly impossible — a $10 credit allotment, a $20 pool at API rates, 400 messages per five hours, and a weekly compute ceiling are four units that don't convert. In April we said "AI cost observability" would become a real category. The June reset just made it mandatory: you can't manage four meters you can't read.

The playbook, updated

What we'd actually do, in order:

  • Know your dates. June 1 already happened. June 15 (Claude Code pools) and June 18 (Gemini CLI shutdown) are still ahead as this publishes.
  • Meter before the vendor does. Log tokens per workflow now, so when the pool shows up you know which cron is the expensive one — and whether it's worth what it burns.
  • Treat promo capacity as borrowed. If your workflow only pencils out on a promotional multiplier or bonus credits, it doesn't pencil out.
  • Tier ruthlessly. Frontier models for the few tasks that need them, cheaper or open models for the bulk. This was advice in April. It's arithmetic in June.
  • Keep one foot portable. Google just deleted a tier people built on. Vendor-agnostic harnesses and exportable workflows are what make a billing email an inconvenience instead of an outage.

The quiet part, again

In April we called the repricing "growing up" — painful short-term, healthier long-term. June is the same story with less ambiguity. The vendors have stopped pretending agents are chatbots. The meters make the cost of autonomy visible, and visible costs are the ones you can engineer against.

The era where a subscription was a flat field of compute is over, for everyone, on every platform, as of this month. What replaces it rewards exactly the builders who know what their agents do, what it costs, and why it's worth it.

Which was always the job anyway.

Sources

Get the next one in your inbox.

We publish when there's something real to send — essays, teardowns, and field notes from a one-engineer workshop running on practical AI. No spam, unsubscribe anytime.

Subscribe →