AI Agent Costs Are Spiraling. Here's How to Stop It.

Why Are AI Agent Costs Suddenly Out of Control?

AI agents cost more than people expect because they don't just answer questions. They plan, retry, call tools, and chain steps together. Every one of those steps burns tokens. A reasoning model like o3 or Claude Opus working through a multi-step task isn't doing one inference call. It's doing dozens, sometimes hundreds. That's the bill that shows up and surprises finance teams.

The Economist reported in June 2025 that companies of all sizes are scrambling to manage spiraling AI costs driven specifically by agentic applications and reasoning-heavy models. This isn't a future problem. It's happening now, in production, at businesses that thought they had a handle on their AI spend.

For SMBs without dedicated ML engineers or FinOps teams, the risk is worse. There's nobody watching the meter.

What Actually Drives Token Costs Up?

Three things cause most of the blowouts we see:

1. Reasoning models used where they aren't needed. Models like o3, o1, or Claude Opus with extended thinking are powerful, but they're priced to match. Using a reasoning model to summarize a support ticket or draft a routine email is like hiring a senior consultant to stuff envelopes. The capability is there, but so is the invoice.

2. Agents that loop without guardrails. Agentic systems retry on failure, re-plan when stuck, and sometimes enter loops that nobody wrote an exit condition for. A workflow meant to process 100 records can silently process 1,000 if something goes sideways and nobody set a hard stop.

3. Long context windows stuffed with unnecessary data. Passing an entire CRM record, a full email thread, or a long document into every agent step when only a slice is needed inflates every call. Context management isn't glamorous, but it's often the single biggest lever on cost.

The issue isn't that AI is expensive. The issue is that most teams have no idea what they're spending per workflow, per task, or per customer interaction.

How Much Are We Actually Talking About?

Let's put real numbers on this. GPT-4o runs at roughly $2.50 per million input tokens and $10 per million output tokens as of mid-2025. OpenAI's o3 model runs significantly higher. Anthropic's Claude Opus 4 is priced in a similar premium tier.

A simple chatbot exchanging a few hundred tokens per conversation is cheap. An agent that pulls context, reasons through options, calls three external tools, and writes a structured output might use 20,000–80,000 tokens per run. Run that 500 times a day and you're looking at real money, fast.

For comparison, a well-scoped GPT-4o mini workflow doing the same functional job as a bloated Opus agent might cost 10–20x less per run with nearly identical output quality for routine tasks. Model selection alone is often a 60–80% cost reduction opportunity.

What Should SMBs Watch For?

Here are the signals that your AI spend is getting away from you:

• No per-workflow cost tracking. If you can only see total API spend, not cost per use case, you can't manage it.
• Reasoning models as defaults. If your team reached for o3 or Opus because it felt safer or smarter, check whether the task actually required it.
• Agents without token budgets or step limits. Any agentic system running in production needs hard caps on retries, steps, and context length.
• Costs growing faster than usage. If your spend doubled but your output only grew 20%, something is running inefficient loops or using the wrong model tier.
• No model routing logic. Sending every request to the same model regardless of complexity is a budget leak. Routing simple tasks to smaller models and complex tasks to larger ones is table-stakes cost management.

What Does a Basic AI Cost Governance Setup Look Like?

You don't need a FinOps team. You need three things in place:

Logging at the workflow level. Every agent or automation should log token input, token output, model used, and task type. LangSmith, Langfuse, and Helicone all do this. Most teams aren't using any of them.

Model tiering by task complexity. Define which task types get which model tier. Routine extraction, classification, and drafting go to GPT-4o mini or Haiku. Complex reasoning, synthesis, or high-stakes decisions go to larger models. Make this a written policy, not a vibe.

Hard caps on agentic workflows. Every agent loop needs a maximum step count and a maximum token budget per run. When it hits the cap, it either escalates to a human or exits gracefully. This single change prevents the worst runaway cost scenarios.

| Tool | Purpose | Free Tier | Pricing Model | |---|---|---|---| | Langfuse | LLM observability and cost tracking | Yes | Usage-based | | Helicone | API proxy with cost analytics | Yes | Usage-based | | LangSmith | LangChain-native tracing and evals | Yes | Usage-based | | OpenMeter | Usage metering for any API | Limited | Usage-based |

Is This Problem Getting Better or Worse?

Worse before better. The entire industry is pushing toward more agentic, more autonomous, more reasoning-heavy systems. That's the direction the big labs are investing in and it's what the demos show at every conference. The capability gains are real. So are the cost curves.

SMBs that build governance now, before they've scaled these workflows, are in a much better position than those trying to retrofit controls after the bills have already shocked the CFO. The businesses getting hurt are the ones who moved fast on deployment and slow on observability.

The goal isn't to avoid AI agents. They deliver real leverage. The goal is to run them like a business operation, not a science experiment.

What We'd Actually Do

• Audit every production AI workflow this week. Pull your API bills, match spend to use cases, and identify which workflows have no per-run cost visibility. That list is your immediate risk surface.
• Implement one observability tool before deploying another agent. Langfuse or Helicone will take an afternoon to set up and will immediately surface which workflows are burning disproportionate tokens.
• Write a one-page model selection policy. Decide which task types get which model tier, set it as team standard, and review it quarterly as model pricing evolves. This is the highest-leverage governance move most SMBs aren't making.