Home

Blog

AI Agents & Orchestration

Brad Griffith

Nov 5, 2026

Token cost is top of mind everywhere these days. There is no getting around it, even when token prices fall. AI bills rise when token prices fall because the number of tokens each task consumes has grown faster than the price per token has dropped. Agents plan, call tools, read results, and loop, and each step re-sends context as input. TNW reported in June 2026 that GPT-4-equivalent performance cost roughly $0.40 per million tokens, down from $20 in late 2022, a 98% reduction, while enterprise AI bills rose an estimated 320%. Price per token measures the meter’s rate. Tokens per outcome is what moved.

This is a volume problem, and volume is controllable.

How large is the gap between price and bill?

The gap is large in the reporting available, and the reporting is secondary. TNW cited an average enterprise AI budget rising from $1.2 million a year in 2024 to $7 million in 2026. It also estimated that a simple linear workflow in 2023 cost about $0.04 per interaction, while an orchestrated agentic system in 2026 costs roughly $1.20, about 30 times more.

Even if we treat these figures simply as directional, the direction, more tokens per task, is consistent with BCG’s analysis of the token meter.

Why do agents use so many more tokens?

Agents use more tokens because one task becomes many model calls. A chat answer is one call. An agent working a task reasons, picks a tool, reads the result, reasons again, and repeats, and every call carries the growing conversation as input.

BCG lists the drivers of a token bill: prompt length, retrieved context, output length, model choice, reasoning effort, tool choice, cache behavior, and the number of loops an agent runs before it stops. Several of those multiply together.

What are the real-world examples?

TNW’s reporting, drawn from TechCrunch and Axios, describes companies surprised by usage. Uber exhausted its 2026 AI coding budget by April, Microsoft revoked Claude Code licenses six months after enabling them, and one company reportedly ran up a $500 million bill in a month after forgetting to set usage limits. Jellyfish reported per-developer consumption up about 18.6 times in nine months.

The common cause in those accounts? Missing limits and missing attribution, not a pricing change.

Does cheaper tokens mean lower cost?

Here’s the thing, cheaper tokens lower cost only when volume stays flat. When a price drop makes new uses viable, volume rises, and when agents replace single calls with loops, volume rises more. The meter’s rate and the meter’s reading are separate variables.

The right target is cost per successful outcome. BCG proposes tracking cost per outcome and a ratio it calls return on AI, so spend is judged against what the work produced.

Which controls bend the curve?

Four controls bend the curve: limits, routing, caching, and execution design. Limits stop runaway loops, routing sends easy tasks to cheaper models, and caching avoids paying twice for repeated input. Execution design reduces how many calls a task needs.

Workato Model Gateway provides rate and token limit policies per team, project, or automation, rule-based routes, and usage reporting by input, output, cache read, and cache write. Enterprise Skills package multi-step work as one governed action.

How do skills reduce tokens per task?

Skills reduce tokens per task by moving orchestration from the model to the platform. When an agent calls raw tools, the model decides each step and pays for it. When it calls a skill, the platform runs the steps and the agent receives one result.

In a Workato-run test on one Jira scenario, with the same account and model, Atlassian’s hosted MCP server needed 3 tool calls, 164,000 tokens, and a 142 KB payload. A Workato server needed 1 call, 31,000 tokens, and 6 KB. The test was vendor-run, covers one scenario, and compares only against Atlassian. Separately, an unnamed Fortune 100 customer running 39 MCP servers reported a 93% cut in LLM token consumption after shifting execution to the platform.

What should finance ask engineering?

Finance should ask five questions. Which workflows use the most tokens? What outcome does each produce? What share of input is cached? Which teams have limits? What happens when a loop does not stop?

BCG offers similar questions for CFOs and CIOs, including a standing review of the highest-token workflows against the outcomes they produce. A gateway that meters every call makes the first three answerable from a dashboard.

Summary: Why AI Bills Rise When Token Prices Fall

AI bills rise because tokens per task grew faster than price per token fell. Agents make many calls per task and re-send context each time, and the reported figures are secondary and directional. Measure cost per outcome, not cost per token, and use limits, routing, caching, and skills to reduce tokens per task.

FAQ

Why are AI costs going up if tokens are cheaper?

More tokens per task. Agent loops multiply calls and context.

How much have token prices fallen?

TNW reported GPT-4-equivalent performance falling from about $20 to about $0.40 per million tokens between late 2022 and mid-2026, a 98% drop.

What is a runaway agent loop?

An agent that keeps calling the model and tools without reaching a stopping condition. Token limits cap the damage.

How can I forecast AI spend?

Meter by workflow and attribute to owners. Forecast cost per outcome times expected volume.

Does routing to smaller models help?

For tasks smaller models handle well, yes. Test quality first.

About the Author

More on AI Agents & Orchestration

Discover the latest product updates and features

A CISO’s tips for running agentic AI in production

Nov 2, 2026

Workato vs Boomi: Which Is Stronger for MCP and Agentic AI?

Oct 20, 2026

The Google Workspace MCP That Teaches Itself