What changed, in one paragraph
For the past year, if you wrote a script that called Claude through the Agent SDK, ran claude -p in a cron job, or wired Claude Code into a GitHub Actions workflow, all of that automation quietly drew on the same flat-rate Pro or Max subscription you pay for. That subsidy ends on 15 June 2026. Anthropic announced the change on 14 May, and from the cut-over date your programmatic and agentic Claude usage is metered against a brand-new, separate monthly credit pool, billed at full API list rates. The pool is small, it does not roll over, and you have to claim it once before it switches on. Your interactive chats and your hands-on-keyboard Claude Code sessions are not affected. Everything you automate is.
- Effective date: 15 June 2026 (announced 14 May 2026).
- Who moves to the credit pool: Claude Agent SDK,
claude -pheadless or print mode, Claude Code GitHub Actions, and third-party agents that authenticate through your subscription. - Who stays on the flat subscription: interactive Claude Code in a terminal or IDE, and web, desktop and mobile chat.
- Credit size: roughly $20 a month on Pro, $100 on Max 5x, $200 on Max 20x, billed at standard API rates, per user, with no rollover.
The credit is opt-in. You must claim it once through your Claude account before it starts refreshing. If you do not claim it, your automation can simply stop on the cut-over date with no warning baked into your pipeline. Check this before 15 June, not after a failed nightly build.
Old flat-rate world versus the new credit model
The practical shift is from an effectively uncapped, heavily subsidised flat fee to a metered pool that behaves like an API budget with a ceiling. The table below lays out what moved where.
| Dimension | Before 15 June (flat subscription) | From 15 June (agent credit pool) |
|---|---|---|
| How agent usage is priced | Drew on flat Pro/Max limits, effectively subsidised | Metered at standard API list rates |
| Monthly ceiling | Soft subscription rate limits | $20 Pro · $100 Max 5x · $200 Max 20x |
| Rollover of unused budget | Not applicable | None — resets each month |
| When the budget runs out | Throttled, then resumed | Automation stops unless overflow billing is on |
| Activation | Automatic with subscription | One-time opt-in claim required |
What counts as agent usage — and what does not
The dividing line is whether a human is in the loop at a keyboard or whether code is calling Claude on your behalf. That distinction decides which pool pays. If you have ever felt fuzzy on where Claude Code ends and the Agent SDK begins, this is the moment it starts costing money to be wrong.
| Usage | Pays from |
|---|---|
| Interactive Claude Code in a terminal or IDE TUI | Flat subscription |
| Web, desktop and mobile Claude chat | Flat subscription |
claude -p "..." headless or print mode |
Agent credit pool |
| Claude Agent SDK in your own script or service | Agent credit pool |
| Claude Code GitHub Actions in CI | Agent credit pool |
| Third-party agents on your subscription (e.g. Zed) | Agent credit pool |
The economics matter because the subsidy was large. Several independent write-ups put the old flat-rate effective discount at roughly fifteen to thirty times cheaper than API pricing for heavy agent users. Stripping that subsidy out is the whole point: Anthropic is steering production automation towards API keys, where usage is metered and observable, and keeping the flat subscription for the interactive experience it was designed for.
Do a five-minute audit of your crontab, your GitHub Actions YAML and any home-grown wrappers. Every place you call claude -p or import the Agent SDK is now metered. The most common surprise is a nightly job or a PR-triggered review action that nobody remembers writing, quietly burning the new credit on every push.
What it costs in practice at API rates
Because the pool now bills at standard API list rates, you can estimate how far $20, $100 or $200 actually stretches. As a reference point, Sonnet 4.6 lists at roughly $3 per million input tokens and $15 per million output tokens; Opus 4.7 at roughly $5 and $25. A modestly sized automated code-review pass on a pull request can run tens of thousands of tokens once you include the diff, the surrounding files and the model's response. Do that on every push across a busy repository and a $20 Pro credit can be gone well before month end.
Two levers move that maths in your favour, and they are the same levers that matter for any API workload:
- Prompt caching. If your agent re-sends a stable system prompt or a fixed slab of repository context on every call, caching that prefix turns repeat reads into a fraction of the fresh-token cost. At a metered pool this is no longer a nice-to-have.
- Model routing. Not every automated step needs Opus. Routing the cheap, high-volume steps — triage, classification, summarising a diff — to Sonnet, and reserving Opus for the hard reasoning, can multiply how many runs your monthly credit buys.
We have a longer walk-through of both techniques in Cut LLM Bills 60%: Prompt Caching and Model Routing, and a market view of where Claude Code sits in the wider tooling landscape in AI Coding Stack 2026: Cursor and Claude Code in Production.
Every article here is written by a Verified Builder. Want your name on the next one?
AI Tech Connect lists AI engineers, founders and researchers across India and the UK — and the people hiring browse it to find them. Adding your profile is free.
Become a Verified Builder →What a builder in India or the UK should actually do this week
This change lands hardest on exactly the people AI Tech Connect is built for: solo builders and small teams running automations on a Pro or Max plan because it was cheap and predictable. An indie shop in Pune wiring Claude into a release pipeline, or a two-person studio in Manchester running an overnight documentation agent, both now have a metered ceiling where they previously had a flat fee. Here is the order of operations.
- Claim the credit now. It is opt-in. Log into your Claude account and switch it on before 15 June so nothing stops silently.
- Audit what you automate. Grep your repositories and cron jobs for
claude -pand Agent SDK imports. Know exactly what is metered. - Decide on overflow billing deliberately. Leaving it off means automation halts when the pool empties — safe, but it can break a release. Turning it on means usage continues at API rates beyond the cap — convenient, but uncapped spend. Pick on purpose, per project.
- Meter the overflow. If you enable overflow, set up your own alerting on spend. The pool gives you a soft warning; your finance function will want a hard one.
- Move shared production automation to API keys. A per-user pool is sized for one person experimenting. A team CI pipeline that several people trigger should run on an API key, where billing is metered, observable and not capped at $20.
- Consider routing across providers. If a step does not specifically need Claude, an OpenRouter-style abstraction or a cheaper open-weight model behind the same interface can absorb the high-volume, low-stakes calls and keep your Claude credit for the work that earns it.
If you bill clients for automation you run on their behalf, this is a contract conversation, not just a config change. Metered API-rate billing is much easier to pass through transparently than a flat subscription was. Move client-facing automation onto its own API key per client, and your cost-of-goods becomes a clean line item.
The bigger signal
Zoom out and this is the same story playing across the tooling market in 2026: the flat-rate, all-you-can-eat era for AI automation is closing, and metered usage is becoming the default. GitHub Copilot moved to usage-based billing earlier in the year; Anthropic is now drawing a firm line between the interactive product you pay a subscription for and the programmatic capacity you meter. For builders, the lesson is not to panic about a single $20 line item. It is to treat agent spend as a real, measurable input cost — the way you already treat compute and storage — and to design your automations so that caching, routing and provider choice keep that cost honest. The teams that internalise this now will quietly out-ship the teams that get a surprise invoice in July.
The official source of truth is Anthropic's help centre. Verify the exact figures for your own plan there before you make budget decisions, since per-seat and enterprise terms differ from the headline Pro and Max numbers.