A few months ago, Azeem Azhar — the writer behind Exponential View — built an AI agent to help him research and write. It was useful. It was also expensive: between $50 and $60 a day, with one wild spike hitting $500 in a single day.
Today, that same agent runs for about $6 a day.
That's not a typo. The cost of running a capable AI agent has dropped by roughly 90 percent in less than a year. And the story of how that happened — and what it means for anyone running a business, a federal program, or a department — is worth understanding, because the implications go far beyond the price tag.
How the floor fell out
The drop came from two forces working together.
First, model routing. Instead of sending every query to the most powerful (and most expensive) model, the agent now decides which model to call based on the complexity of the task. Simple lookups go to cheap models. Hard reasoning problems get the expensive ones. The agent itself handles the triage.
Second, the Anthropic-OpenAI price war. Over the past year, both companies have cut API prices repeatedly as they compete for developer mindshare. When the two market leaders drop prices, everyone else follows. The result is a market where capable AI is genuinely cheap.
Ramp, the corporate spend management platform, tracks what companies actually pay for AI. Their data shows the top 1 percent of firms spent a median of $7,400 per employee on AI in July. The median firm? $11.95 per employee.
That spread — $7,400 versus $12 — is itself a story. Some organizations are investing at a level that suggests they've found real leverage. Most are still dipping a toe in.
The hidden problem: cheap agents hide expensive sprawl
Here's where it gets tricky. When an agent costs $6 a day, it's easy to spin up dozens of them. A few here for customer support, a few there for data analysis, a handful for internal research. Each one is cheap. Nobody approves a line item for $6.
But those agents don't exist in isolation. They call APIs. They store data. They generate outputs that need review. They create dependencies. And when the count hits 50 or 100 or 500 agents, nobody has a consolidated view of what they're costing — not just in API fees, but in the human time required to supervise, maintain, and trust their outputs.
This is the cost-observability problem, and it may be the next operational bottleneck for organizations that adopt AI aggressively.
A cautionary tale in plain numbers
Amazon once ran a Claude-based project that cost $1.8 million over five months before anyone noticed. Not because the tool didn't work. Because nobody was watching the bill.
For a company of Amazon's scale, $1.8 million is rounding error. But the pattern — an AI project running for months without spend visibility — repeats at every level. The only difference is the dollar amount. At a small business or a mid-size federal contractor, the same blind spot could produce a very unwelcome surprise.
What spend observability looks like in practice
The organizations that avoid this trap share a few habits:
- Agent-level cost tracking. They know what each individual agent costs per day, per week, per month. Not a general "AI spend" line item. Per-agent granularity.
- Human-time cost included. The $6/day agent costs $6 in API fees and potentially $60 or $600 in the time someone spends verifying its outputs. Good shops track both.
- Off-ramp triggers. They define, before deploying an agent, what would cause them to turn it off. A cost threshold, a accuracy rate, a response-time floor. These are set in advance, not retroactively.
The macro question nobody is asking
Here's the question that keeps coming back to me: If agent costs have dropped 90 percent in a year, and deployment is accelerating, and spend observability is still catching up — where does all that money go?
Azeem Azhar ends his piece with a pointed question: "Has $494 a day just disappeared from genAI revenue?"
It's not rhetorical. When costs drop that fast, the savings have to land somewhere. Either they flow back to the customer (cheaper prices), or they flow to the provider (wider margins), or they disappear into unnoticed expansion — more agents, more usage, more hidden cost.
For anyone running an organization right now, which of those three outcomes is happening on their watch is worth knowing.
What you can do this week
Pull last month's AI spend — all of it, not just the line items you know about. If you can't get a per-agent or per-use-case breakdown, that's your first finding. The cost-observability gap is real, and the first step is measuring what you're spending. The second step is deciding whether the invisible $494-a-day pattern is working for you or against you.