Anthropic released Claude Opus 5.5 on September 22, saying the model performs near its higher-end Fable 5.1 level on most work while costing 40% less to run than Opus 5. The company also raised five-hour usage limits for Pro, Max, and Team subscribers and gave users a saveable rate-limit reset, according to Techmeme’s summary of ZDNet’s report.
Alongside the model improvements, the launch highlights an operating constraint: usage limits can interrupt work the business depends on.
AI access is becoming workflow infrastructure
Seat count, monthly cost, and included models are useful procurement details. Add the expected workload: when people need access, how long tasks run, and what happens if a quota stops them midway.
Compare a short email draft with a project lead’s session reviewing meeting transcripts and updating a timeline. The longer session may reach a quota sooner, depending on context and tool use. Test those workloads on the plan the team actually has.
During the trial, record interruptions and workarounds. Switching tools or accounts may mean reloading context and checking access again. Include that time when judging whether the plan supports the work.
Separate run cost from available capacity
Anthropic’s stated 40% run-cost reduction and its higher subscriber limits address different constraints. Lower unit cost affects the bill; available capacity affects whether the team can finish at the time the work is needed. A proposal review forty minutes before a client call has less room to wait for a reset than a background research task.
OpenAI’s Enterprise Signals data from August showed agent-style work accounting for 64% of combined Codex and ChatGPT output among its enterprise customers as of June 2026. That is an output-share measure for those products and customers. It does not establish the same usage pattern across other providers.
For delegated work, estimate task duration, context size, and bursts of activity. Operations and procurement need that workload estimate when choosing capacity.
Treat limits like a process design question
List roles expected to use AI for sustained work, then estimate their task frequency and duration. In a hypothetical capacity plan, ten casual users and two power users may need different provision from twelve casual users.
Identify handoffs that cannot wait for a reset. If an AI assistant cuts off during a research sprint, proposal review, live implementation session, or end-of-day client summary, its quota creates an unreliable workflow. Record the delay and the fallback the team used.
Give access design to an owner who can review personal upgrades, tool switching, and heavy work routed through a colleague’s higher plan. Those workarounds need permission and continuity checks as well as a cost review.
Choose capacity to fit the work. Options include role-based access, a shared workflow that sends heavy tasks through a smaller number of trained operators, a usage policy defining appropriate tasks, or vendor negotiation. A larger plan is one option among those choices.
Observe a week of actual use
If your AI program is still in the license-counting phase, this is the next audit to run. Pick the roles where AI is supposed to change the work. Watch how they actually use it for a week. Note where limits, resets, plan tiers, permissions, or tool switching interrupt them.
Use the interruption record to choose a plan change, narrower task scope, or an approved fallback. Check the same workflows again after the change to see whether people can complete the work when it is needed.