We tracked 200K AI requests. Here's where the money actually goes

#ai #productivity #devtools #openai

Six months ago I posted here about CostLens — a tool to reduce OpenAI costs. Since then, we've completely rebuilt it based on one question a VP asked our team:

"How much faster are we delivering with AI? What's the number?"

Nobody could answer. We had cost dashboards, but no way to connect spend to output. So we built that.

What CostLens does now

It's not just cost tracking anymore. It's three things:

Cost attribution — not "you spent $4,200 on OpenAI" but "code review costs $340/mo, customer support costs $89/mo, and classification costs $12/mo." Per-feature, per-developer.
Smart routing — simple prompts automatically go to cheaper models. No code changes. We're seeing 30-40% savings on teams that were sending everything to GPT-5.4.
Productivity tracking via MCP — works inside Claude Code, Kiro, and Cursor. Tracks sessions, commits, PRs, and time-to-ship. Generates weekly reports your VP can forward to justify the AI budget.

The setup is still one line

import { CostLens } from 'costlens';
const costlens = new CostLens({ apiKey: 'cl_...' });
const openai = costlens.wrapOpenAI(new OpenAI());
// Done. Every request is tracked and optimized.

For coding agents (Claude Code, Kiro, Cursor):

{
  "mcpServers": {
    "costlens": {
      "command": "npx",
      "args": ["-y", "@costlens/mcp-server"],
      "env": { "COSTLENS_MCP_KEY": "your_key" }
    }
  }
}

What we learned from tracking 200K requests

52% of requests were simple tasks running on expensive models (classification, extraction, yes/no questions)
Smart routing cut our own bill from $4,100 to $2,337/month
The "code review" feature was 3x more expensive than we thought
Sessions with no output (abandoned/looping) accounted for 15% of spend

Kill switch

One thing we added after a background agent burned $50 in a weekend: a Slack kill switch. When spend velocity spikes, you get a message with a "Pause" button. One click stops that specific agent session. Auto-
resumes after 30 min.