Anthropic splits Claude billing into separate pools as AI agents strain subscription model

Anthropic Introduces Dedicated Credits for AI Agent Usage

Anthropic is fundamentally restructuring how its Claude subscriptions handle AI agents and programmatic automation, announcing that starting June 15, paid subscribers will receive separate monthly credits specifically for agentic tool use 1

1

. The move represents a significant shift in how AI companies manage the tension between flat-rate subscriptions and resource-intensive autonomous agents that can burn through computing power at unprecedented rates.

Under the new system, Pro subscribers paying $20 per month will receive a dedicated $20 credit for programmatic usage, while Max 5x users get $100 and Max 20x subscribers receive $200 in monthly credits 2

2

. Team Standard plans will allocate $20 per seat, and Team Premium plans will provide $100 per seat 5

5

. Critically, these credits do not roll over between billing cycles and will be charged at API rates once exhausted.

The OpenClaw Problem That Forced Change

The restructuring effectively reverses Anthropic's controversial April 2026 decision to ban third-party agent frameworks like OpenClaw from using Claude subscriptions 4

4

. Boris Cherny, Head of Claude Code, explained at the time that third-party services were "really hard for us to do sustainably" because they bypassed caching mechanisms that allow Anthropic to offer flat-rate subscriptions . Some Claude subscribers were paying $20 to $200 monthly but consuming hundreds or even thousands of dollars worth of tokens through autonomous agents, creating an unsustainable financial position for the company.

Interactive Versus Programmatic Usage Split

The new billing structure creates a clear divide between interactive and programmatic workflows. Interactive use—typing prompts in Claude Code, Claude Cowork, or Claude.ai and receiving responses with a human in the loop—continues to draw from standard subscription limits 1

1

. However, programmatic interaction through Anthropic's Agent SDK, headless mode (claude-p), GitHub Actions, or third-party tools like OpenClaw now counts against the separate credit pool 4

4

.

Once the programmatic credit is exhausted, users who have enabled "extra usage" billing will be charged at standard API rates, while those without this feature enabled will face service cutoffs 2

2

. The non-rollover nature of credits means users must calibrate their programmatic usage carefully to extract full value from their subscriptions.

Industry-Wide Shift Away from Unlimited AI

Anthropic's move reflects a broader industry reckoning with how autonomous software consumes computing resources compared to humans. While a human might send dozens or hundreds of prompts daily, autonomous coding agents can generate thousands of requests, run continuous tests, browse the web, and recursively call models 3

3

. ServiceNow and Uber have already burned through their AI token budgets for the entire year, according to The Information 3

3

.

Microsoft's GitHub Copilot is pursuing a similar transition to metered credit allotments 1

1

, while infrastructure providers like Fireworks AI and Together Computer already price serverless inference by token usage rather than flat subscriptions 4

4

. The pattern suggests that "all-you-can-eat" AI subscriptions may not survive the agent era.

User Backlash and Competitive Response

The announcement drew sharp criticism from developers. Claude Code Product Manager Noah Zweben's X post about the new rules attracted replies calling the changes "gaslighting" and threatening to switch to competing services 3

3

. Users on social media characterized the move as "a downgrade dressed up as a feature," noting that power users could hit the cap within days while others might waste unused credits in months with lower programmatic demand 4

4

.

Meanwhile, OpenAI is taking the opposite approach in the short term, with Sam Altman announcing that the company is giving new business customers two months of free usage 3

3

. However, industry observers note that everyone faces the same economics and will eventually need to move away from unlimited use. For enterprises, the shift signals that production agents running continuously with long workflows will increasingly require rationed environments with budget controls and predictable limits 4

4

.