Token Tsunamis: why agentic AI systems blow up your costs

Redaktion · · 4 Min. Lesezeit

Several industry and research analyses from the first half of 2026 now put concrete numbers on what agentic AI systems cost to run — and those numbers explain why budgets keep breaking. The core finding: an autonomous agent that works through tasks in loops consumes a multiple of the tokens a simple chat prompt uses. The analyses circulate the term “token tsunamis” for this — cost waves created by loops, retries, and constant context reloading. The figures below come from industry and research sources and should be read as estimates, not as hard vendor data.

What was true before

In the chat era, the token math was simple: one prompt in, one answer out. Costs could be estimated roughly from average prompt and response length. Planning a feature meant multiplying expected requests by average token count by price per token — and landing close to reality.

That model no longer holds for agentic systems. An agent is not a single request but a chain: it reads context, plans, calls tools, checks results, corrects itself, and starts over on errors. Each of those steps is its own model call with its own token consumption. What looks like “one task” from the outside is often internally a two- or three-digit number of calls.

What is true now

1. Loops are the dominant cost driver, not prompt length. The analyses show that what matters is not the size of a single prompt but how often the agent iterates. An agent that runs ten correction rounds on a hard problem can consume roughly 50 times a clean pass. The gap between a well-behaved agent and one stuck in loops is therefore not 20% but one to two orders of magnitude.

2. Context reloads add up silently. Every time an agent reloads its accumulated context into the prompt, you pay for those tokens again. On long runs with many tool results, context reloading becomes its own cost block — one that is entirely missing from list-price planning.

3. Hierarchical architecture cuts cost measurably. One solution named in the analyses: not every model in the system needs to be a frontier model. If cheap budget models take the worker roles and only the orchestrating lead agent uses a frontier model, one source reports reaching about 97.7% of full-frontier accuracy at roughly 61% of the cost. It is not a free win, but it is a clear lever.

Context

The numbers should be read with caution — they come from vendor-adjacent analyses and individual research papers, not a standardized benchmark. The order of magnitude, however, is consistent across several independent sources: agentic systems do not cost a little more per task but a multiple. That matches what large adopters report — for example Uber, which spent its 2026 AI coding tool budget in four months.

For practice this means: multi-agent orchestration is not a pure engineering topic but also a budget question. Anyone running several agents in parallel without measuring and capping consumption builds in a cost risk that, in the worst case, sits two orders of magnitude above plan. The answer is not to avoid agentic tools — they are useful — but to treat them like a variable cost line.

What you can do now

If you build multi-agent systems: set a token budget per task and abort runs that exceed it. An agent without a loop limit is an open cost risk.

If you design the architecture: check which roles really need a frontier model. Worker steps (reading files, checking formats, simple tool calls) often run well enough on cheaper models — keep the frontier model for the orchestrator.

If you own consumption: measure tokens per typical task in a slice test before extrapolating. List prices say nothing about your agents’ real iteration count — and that is exactly what decides the bill.

Entdecke mehr