If you run Claude in a CI pipeline, an agent loop, or any script that calls claude -p, your monthly budget just got its own rules. Starting June 15, Anthropic separated all programmatic Claude usage from your subscription's shared pool and moved it to a standalone credit billed at standard API rates. That $20 monthly credit sounds reasonable until you run the math on a single Opus session.
What exactly changed on June 15?
Programmatic usage now draws from a separate credit pool, not the same bucket as your interactive chat and terminal Claude Code sessions. Until June 14, every call you made through the Agent SDK, claude -p, or a GitHub Actions workflow drew from the same allowance as your browser-based conversations. As of today, those two buckets are completely independent.
Your interactive work is untouched. Claude Code from your terminal, conversations on claude.ai, IDE co-coding: all of that runs exactly as before. The split only affects automated, unattended usage.
Here's what falls under the new programmatic credit:
- The Claude Agent SDK
- The
claude -pCLI command (headless, non-interactive) - Claude Code GitHub Actions
- Third-party apps built on the Agent SDK
Anthropic sent the announcement by email on June 8, exactly one week before the change took effect. If you missed it, you're discovering the new rules right now.
How much credit does your plan include?
Every paid plan gets a fixed monthly credit for programmatic use, billed at Anthropic's standard API rates. The credit is included in your plan; you don't pay anything extra on top of your subscription fee.
| Plan | Programmatic credit |
|---|---|
| Pro ($20/month) | $20 |
| Max 5x ($100/month) | $100 |
| Max 20x ($200/month) | $200 |
| Team (Standard) | $20 per seat |
Unused credits don't roll over. They reset with each billing cycle, no exceptions.
What does a session actually cost?
It depends entirely on which model you run. A medium-weight code session, roughly 200,000 input tokens and 50,000 output tokens, runs about $1.35 on Sonnet 4.6. On Opus 4.6, that same session costs $6.75. Five times as much for the same number of tokens.
| Model | Input (per M tokens) | Output (per M tokens) | Cost per session* |
|---|---|---|---|
| Claude Sonnet 4.6 | $3 | $15 | ~$1.35 |
| Claude Opus 4.6 | $15 | $75 | ~$6.75 |
*Estimated for a session with 200,000 input and 50,000 output tokens. See current model prices in the model tracker.
On Sonnet, the Pro credit covers roughly 15 sessions a month. On Opus, it covers three. For occasional use that's fine. A developer who automates daily PR reviews or refactors through claude -p can burn through it in a single workweek.
For lighter tasks the math is friendlier. A single PR review on Sonnet 4.6, at around 30,000 input and 5,000 output tokens, runs about $0.17. You get over 100 of those in a $20 monthly credit. The pinch comes when sessions get heavy or you default to Opus for everything.
What happens when the credit runs out?
Your automations stop. There's no graceful degradation or warning mid-pipeline. They stop until the credit resets next month, unless you've enabled overflow billing.
Anthropic calls the overflow option Usage Credits. With it enabled, you keep running at standard API rates with no ceiling. You can toggle it on or off at any time in your account settings.
For context: developers tracking this change calculated that effective pricing increases range from 12x to 175x depending on workload type, compared to what heavy programmatic users were implicitly paying under the flat-rate model. That's a wide spread, but the upper end applies to teams running frequent, long-context Opus sessions that previously had no per-token cost at all.
Think of it like the shift from fixed-price shared hosting to pay-as-you-go cloud billing a decade ago. You pay for what you use, but you need to know how much that is first. Most Claude subscription users didn't need to track their token consumption before. They do now.
Why is Anthropic making this change?
The flat-rate subscription model stopped making financial sense once developers started routing heavy agent workloads through it. A Pro user running dozens of agentic sessions per day costs Anthropic a multiple of the $20 monthly fee in compute. The split brings the pricing model in line with actual consumption.
"This removes an affordable path to running production-grade agentic workloads and creates budget unpredictability, because usage is now directly tied to token consumption."
Senior reliability engineer, responding to the June 8 announcement
OpenAI has always charged Codex and the Assistants API on a per-token basis. Google does the same for Gemini API usage. Claude subscription users were used to a model where one flat rate covered automated workloads alongside interactive use. As of June 15, that no longer applies to programmatic calls.
How does this affect your team?
If your team runs Claude Code or the Agent SDK in automated workflows, this is a real budget line item to account for. Consider a five-person engineering team where each developer runs ten Sonnet sessions per week through the Agent SDK. That's 200 sessions a month, at roughly $1.35 each, totaling $270. A Team plan at $20 per seat gives $100 in combined programmatic credit. The remaining $170 comes out of overflow billing at standard API rates.
For larger teams or heavier Opus workloads, the monthly costs scale accordingly. A 20-person team running mixed model workflows can see programmatic bills in the low thousands per month, with no equivalent spending under the previous flat-rate model. Worth noting: credits are per account, not poolable across seats. If you run shared automation on a single account, that account's credit carries the entire load. Distribute automated workflows across team members where that's feasible.
What should you do today?
Start with an audit of your automation footprint. Three concrete steps:
- List your automated Claude usage. Anything running via
claude -p, the Agent SDK, or GitHub Actions now draws from the programmatic credit. Identify every process and which model it uses. - Run the cost math for your workload. Use the session estimates above as a starting point. If you're comfortably under your plan's credit limit, nothing changes in practice. If you're over, decide between upgrading your plan, enabling overflow billing, or moving lighter workflows to Sonnet.
- Audit your model choices. Sonnet 4.6 costs one fifth of Opus 4.6 per session. For most automated tasks, including PR reviews, linting, code summarization, and documentation generation, Sonnet handles the job without the Opus price tag. Reserve Opus for sessions where you genuinely need the extra reasoning depth.
If you're still running Opus in CI pipelines, this is a good time to benchmark Sonnet 4.6 against those specific workflows before overflow billing starts accruing. The TheAIDaily model tracker has current pricing and context window specs for every major Claude version if you want to compare before switching.
