The Claude Billing Split: What Actually Changed on June 15

On June 15, 2026, Anthropic split Claude's billing into two separate consumption pools. If you missed the announcement, you are not alone — it was buried in a support article. But the implications for every engineering team running Claude are significant and immediate.

Here is what changed, what it actually costs you, and what to do about it.

What Actually Happened on June 15

Anthropic announced a structural change to how Claude subscriptions are billed. Starting June 15, 2026, two distinct consumption pools now exist for any Claude subscriber:

Pool 1 — Interactive use: Claude Code, Claude Cowork, and direct Claude web/app interactions. This stays on your plan's usage limit. Your Pro $20/month still gets you the same allocation of Claude. That part didn't change.

Pool 2 — Agent SDK and programmatic access: Claude Agent SDK, claude-p command-line tool, Claude Code GitHub Actions, and any third-party application built on the Agent SDK. This moved to a separate credit pool, sized by plan tier:

Credits refresh monthly. Unused credits do not roll over. When the credit is exhausted, usage moves to standard API rates — but only if you have enabled usage credits. If you have not, Agent SDK requests stop when the credit runs out.

The April 2026 Change You Probably Missed

The June 15 split was announced publicly, but a quieter change shipped in April alongside Claude Opus 4.7: a new tokenizer that fundamentally alters how many tokens Anthropic counts per input.

Under the new tokenizer, the same prompt that registered as 1,000 tokens under Claude Opus 4 now registers as 1,350 tokens under Opus 4.7. The per-token price is technically unchanged — $5 per million input tokens, $25 per million output tokens. But the bill for identical content just went up by as much as 35%.

This is not a pricing increase. It is a tokenizer artifact that nobody announced with a press release. And it has real implications for any team running high-volume workloads:

• Long system prompts: Architectural context, multi-role agent setups, detailed instruction files all tokenize differently under the new tokenizer. The delta is proportional to prompt complexity.
• Code-heavy inputs: The new tokenizer is more efficient for natural language but less efficient for syntax-heavy content. If you are using Claude for code generation or code review, your token inflation will be higher than the average.
• Structured outputs: JSON schemas, XML-formatted responses, and any output with heavy formatting characters see disproportionate token inflation under the new tokenizer.

Combined Effect: What This Costs a Real Engineering Team

Let's take a concrete example. A team of 15 engineers, all running Claude Code in their local environment and via CI, plus one shared production automation pipeline running on the Agent SDK.

Under the old model (pre-June 15, pre-April tokenizer change):

• 15 × Pro ($20/mo): $300/month for interactive use
• Production pipeline: counted against plan limits
• Token inflation: none (old tokenizer)

Under the new model (post-June 15, post-April tokenizer):

• 15 × Pro: $300/month — interactive use unchanged
• Agent SDK credit per Pro seat: $20/month — total $300 for the team
• Production pipeline: burns through $300 in weeks, then hits standard API rates at $5/MTok input
• Token inflation: +35% on every call — so the $300 credit buys ~285M tokens instead of 400M tokens

The real monthly bill for a team running Claude Code locally and Agent SDK in production: $800–$1,200/month where it used to be $300. For the same engineers, doing the same work, with the same tool.

Why This Matters Beyond the Bill

The billing split reveals something important about how Anthropic is positioning its pricing going forward. The company is moving deliberately toward a model where interactive human use and programmatic machine use are priced, metered, and governed independently.

This mirrors a broader industry pattern. GitHub moved Copilot to token-based billing in June 2025. OpenAI has been meter-based from the start. Google has been gradually shifting Gemini pricing toward usage-based for enterprise. Every major AI provider is converging on the same model: access cost + consumption cost, separated.

The implication for engineering teams is not just "your bill went up." The implication is that AI tool spend is now a variable cost that responds to how your engineers use the tools — not just how many seats you have.

What Teams Running Agentic Workflows Need to Do

1. Audit your current Agent SDK consumption. If you are running OpenClaw, Conductor, Claude Code in CI, or any automated agent pipeline on Pro or Max plans — measure your actual token consumption before June 15. The $20 credit per Pro seat will run out faster than you think. A team of 10 engineers running moderate agentic workloads will burn through $200 in less than a week.

2. Move shared production automation to a proper API key. Anthropic's own guidance is explicit: "Teams running shared production automation should use Claude Platform with an API key for predictable pay-as-you-go billing." The Agent SDK credit is sized for individual experimentation, not production workloads. If your pipeline is business-critical, it needs a proper API billing relationship.

3. Benchmark your tokenizer delta. Before assuming cost parity between Opus 4 and Opus 4.7, run your production prompts through Anthropic's tokenizer. The 35% inflation is real for structured outputs and code-heavy inputs. Update your cost models accordingly.

4. Enable budget controls before you need them. When the credit runs out, Agent SDK requests stop. If your CI pipeline is running Claude Code on a Pro seat credit and it hits the ceiling mid-build, the job fails. Set up spend alerts and hard caps before the credit runs out — not after.

How Trimio Handles This

When Anthropic shipped the tokenizer change in April and the billing split in June, most teams found out from their invoice. Trimio customers found out from their dashboard — and had 30 days to adjust before the changes went live.