Claude Code's June 15 billing change was paused. Here's the real math.

The 60-second version

Half of LinkedIn spent this week sharing the same screenshot: "Anthropic ended the Claude Code free ride on June 15 — automated usage now bills from a separate budget." I run Claude Code and agents in production, so I went to the source instead of the screenshot.

Anthropic announced the change. Then it paused it — the same day it was meant to ship. As of today (June 17, 2026), nothing has changed. But the economics that triggered the change are still sitting there, and they will come back. So the useful question is not "what changed" (nothing) but "what would it cost me when it returns" — and that you can compute today.

What was announced, and what actually happened

The announced change (around June 2, effective June 15): "non-interactive" Claude usage would stop drawing from your subscription's usage limits and move to a separate monthly credit, then standard API rates once the credit runs dry.

In scope (non-interactive):

• The Claude Agent SDK in your own projects
• claude -p (Claude Code's headless, non-interactive mode)
• The Claude Code GitHub Actions integration
• Third-party apps that authenticate with your subscription through the Agent SDK

Out of scope: interactive use — you typing in Claude Code or in claude.ai. Never affected.

The announced credit, per plan, monthly, no rollover: $20 on Pro, $100 on Max 5x, $200 on Max 20x.

What actually happened: on June 15 Anthropic posted, word for word, "We're pausing the changes to Claude Agent SDK usage described below. For now, nothing has changed: Claude Agent SDK, claude -p, and third-party app usage still draw from your subscription's usage limits." They added that they will rework it and "share it before anything takes effect."

So if you saw the "the deal is over" post: it described the announcement and missed the reversal. Worth knowing before you forward it to your team.

What the data shows

I built a small cost model at Anthropic's standard API list rates (Sonnet 4.6 at $3 / $15 per million input/output tokens; Opus 4.8 at $5 / $25). One "substantial" Claude Code task is roughly 150K input + 30K output tokens billed. Full assumptions and the script are linked at the end.

Cost per substantial automated task:

• Sonnet 4.6: ~$0.58 with prompt caching, ~$0.90 without
• Opus 4.8: ~$0.97 with caching, ~$1.50 without

What the announced (now-paused) monthly credit would have bought, in substantial tasks/month:

• Pro ($20): ~24 tasks (range 13–34)
• Max 5x ($100): ~124 tasks (66–170)
• Max 20x ($200): ~249 tasks (133–341)

What automation actually costs per month at API rates:

• Interactive only (you typing): $0 — out of scope
• Light automation (2 headless runs/day): ~$35
• PR-review bot (20 PRs/day): ~$480
• Cron agent every 15 minutes: ~$588

The one number that matters: a PR-review bot on 20 pull requests a day is ~$480/month at API rates. Against a $20 Pro credit, that is a 24x gap between what you pay on a flat subscription and what the same work costs metered. That gap — reported elsewhere as a "15–30x subsidy" — is exactly what Anthropic tried to close, and the reason it will try again.

So the viral post's one correct instinct ("a single heavy session can chew through your budget in a few runs") is directionally right for the smallest plan and load-bearing automation. It was just describing a change that isn't live.

Why this matters for production (even while it's paused)

• Your subscription is currently subsidizing automation by roughly an order of magnitude. That is great — and it is not a contract. Plan as if the meter is coming back.
• The risk is not the price. It is the surprise. A flat subscription hides automated consumption until the model changes; then it lands as an invoice instead of a dashboard line.
• Interactive and headless usage have completely different economics. Treating them as one bucket is why teams get blindsided.
• "Set and forget" cron agents are the silent budget killer: 96 runs/day is ~$588/month before a human looks at it.

What to do — by situation

• You only type in Claude Code / claude.ai: do nothing. You were never in scope.
• You run light automation (a few headless runs/day): no action needed now; even the $20 Pro credit would have covered you. Keep it on the subscription.
• You run load-bearing automation (CI bots, Agent SDK in prod, cron agents): meter it now. Split interactive from headless, put a number on the headless half, and decide your billing posture before the change returns — not after.
• You are about to put an agent in production: start it on explicit API / pay-as-you-go billing with a spend cap and a dashboard from day one. Subscriptions are for humans; production automation belongs on metered billing you can see.

How I built the cost model

Inputs are public list prices plus labelled assumptions — nothing fabricated:

• API rates: Sonnet 4.6 $3/$15, Opus 4.8 $5/$25 per million tokens (Anthropic pricing).
• Task size: 150K input + 30K output tokens per substantial task; a lighter recurring task at 60K + 10K.
• Caching: cache reads bill ~0.1x of input; I assume ~70% of input is cache-read on a long session (a 0.3x blended input multiplier).
• 30 days per month.

Cost per task = (input tokens / 1M x input rate x cache multiplier) + (output tokens / 1M x output rate). Credits-to-tasks = credit / cost per task. Scenario cost = runs per month x cost per run. The script prints the table and writes a CSV + JSON.

What this analysis does NOT measure

• List rates only — negotiated or enterprise pricing differs.
• Token-per-task is an assumption, not a measurement of your workload. Your real numbers can be 2–3x in either direction.
• Caching savings vary enormously with how your prompts are structured.
• It ignores the Batch API (-50%), volume tiers, and free code-execution hours.
• These are order-of-magnitude scenarios, not a quote.

The bottom line

Nothing changed on June 15. But "nothing changed" is a snapshot, not a strategy. The subscription subsidy on automated Claude usage is real, large, and explicitly under review by the company that pays for it. If your Claude bill is a flat subscription today and you cannot put a number on what your automation would cost at API rates, that is the gap to close before the change comes back — calmly now, not in a panic later.

Next step

A 15-minute diagnostic applies this same metering to your actual stack — interactive vs headless, what your automation would cost at API rates, where to put the spend caps — for 97€ at godi.ai/audit-ia-quick. Questions: [email protected].