Run a Claude Code capacity check before the 125-unit weekly limit

Claude Code operators on Pro, Max, Team, and seat-based Enterprise plans had 17 percent less weekly capacity on September 14 than the day before, and the change worked out to the same reduction in weekly Claude Code limits compared with the then-current level. Anthropic ended the temporary 50 percent usage boost and set the new permanent weekly limit at 125 units, 25 percent above the original 100-unit baseline and 17 percent below the 150-unit boosted level.

The triage follows from that line: 150-unit workflows have the least slack, 100-unit workflows have more room, and 125-unit workflows sit at the line.

The cut is a capacity audit

The change is a ceiling. Long sessions, retries, and parallel agents now consume the same weekly meter with less headroom. A workflow that ran comfortably under the boost can hit the ceiling earlier in the week. The audit is simple: find the seats near the top, list the workflows that assumed the higher level, and decide which work can move.

Run the 100/150/125 triage

Compare each seat's weekly usage against the new ceiling before the next sprint. Seats that were comfortable under the boost are the first ones to throttle. Use this triage:

• Pull the weekly usage percentage for each seat and note who is near the ceiling.
• List workflows sized to the 150-unit boosted level.
• Mark each workflow as real-time or background.
• Flag background or scheduled work for the Batch API, which has no weekly cap for workloads that do not need real-time responses.
• Flag lower-priority work for cheaper models or a queue.

The triage gives a routing decision. Seats near the ceiling need a route change now. Work that can wait should leave the weekly meter. Work that cannot should keep a clear path to the primary model. Do not treat every slowdown as a weekly-capacity failure. The five-hour rolling window limits did not change.

Move work that does not need speed

Long sessions burn the meter faster. One agent command can make 8 to 12 internal API calls, and a request in an active session can reach 200,000 input tokens by the 15th command. A workflow that looked fine under the boost can break at the new level. A short command is not the same as a long agent run.

Route the work by response need. Real-time means a person is waiting, a build is blocked, or a customer-facing action depends on the answer. Background means the output can arrive later without changing the user experience. Real-time coding help stays on the primary model. Background summaries, scheduled reports, and bulk refactors can move to the Batch API or a cheaper model. If a job can wait, it should not consume the weekly ceiling.

Use a cheaper model for classification, summarization, or routine refactoring. Do not use it for tasks that need deep codebase context or multi-step reasoning. Use a queue when demand is spiky. It holds work until the weekly meter has room. Make the routing rule explicit in the workflow definition. Run the triage now, and keep the real-time work on the model that can answer immediately.

Keep the real-time path clean

Protect the work that users feel. If a developer is waiting for an answer, the workflow should have a clear path to the primary model. If a background job is competing for the same weekly meter, it should move. The core routing rule is simple: speed-sensitive work stays, delay-tolerant work moves.

Review the mix after each major release. New agent features can change how much work a command does. A workflow that was safe before may need a new route now. Assign an owner to the weekly usage review. Keep the route visible to the team.