Free tool · #12

Agent Cost Decomposer

Paste a multi-step agent trace and see the cost of every step — by model, by driver, and any retry loop burning through it.

Prices verified 15 Jul 2026 · 131 models

Trace JSON

{
  "steps": [
    {
      "label": "plan",
      "model": "claude-sonnet-5",
      "input_tokens": 1200,
      "output_tokens": 320
    },
    {
      "label": "route_task",
      "model": "acme-internal-router-v1",
      "input_tokens": 300,
      "output_tokens": 40
    },
    {
      "label": "search_web",
      "model": "claude-haiku-4-5",
      "input_tokens": 480,
      "output_tokens": 140
    },
    {
      "label": "extract_data",
      "model": "claude-haiku-4-5",
      "input_tokens": 860,
      "output_tokens": 95
    },
    {
      "label": "extract_data",
      "model": "claude-haiku-4-5",
      "input_tokens": 845,
      "output_tokens": 90
    },
    {
      "label": "extract_data",
      "model": "claude-haiku-4-5",
      "input_tokens": 872,
      "output_tokens": 98
    },
    {
      "label": "verify_result",
      "model": "gpt-4o-mini",
      "input_tokens": 650,
      "output_tokens": 60
    },
    {
      "label": "summarize",
      "model": "claude-sonnet-5",
      "input_tokens": 2100,
      "output_tokens": 540,
      "cached_input_tokens": 1200
    }
  ]
}

Accepts { steps: [...] }, a bare array of steps, or an OpenAI-style array of { model, usage } entries. The loaded example is an illustrative example, not real trace data.

Cost decomposition

$0.02priced across 7 steps

4 models · 1 step unpriced

Steps 4–6 look like a retry loop on claude-haiku-4-5 — $0.0040 (19% of the priced cost).

Cost per step

1planclaude-sonnet-51,200 in · 320 out$0.0056
2route_taskacme-internal-router-v1unknown300 in · 40 out—
3search_webclaude-haiku-4-5480 in · 140 out$0.0012
4extract_dataclaude-haiku-4-5860 in · 95 out$0.0013
5extract_dataclaude-haiku-4-5845 in · 90 out$0.0013
6extract_dataclaude-haiku-4-5872 in · 98 out$0.0014
7verify_resultgpt-4o-mini650 in · 60 out$0.00013
8summarizeclaude-sonnet-52,100 in (+1,200 cached) · 540 out$0.0098

By model

Model	Steps	Cost	Share
claude-sonnet-5	2	$0.02	74%
claude-haiku-4-5	4	$0.0052	25%
gpt-4o-mini	1	$0.00013	1%
acme-internal-router-v1	0 / 1	$0	0%

Biggest drivers

Step 8 — summarize claude-sonnet-5$0.0098 (47%)
Step 1 — plan claude-sonnet-5$0.0056 (27%)
Step 6 — extract_data claude-haiku-4-5$0.0014 (7%)

This is what Telemeter does for every agent run, automatically.

Use this tool from code — or hand it to your agentagent prompt · REST · MCP

Hand this tool to your coding agent

Paste into Claude Code, Cursor, Codex, or any agent. It learns to call this tool over REST or MCP, reaches for it when you ask, and keeps our sourced/pending labels honest.

You can use telemeter.ai's free "Agent Cost Decomposer" — a keyless, read-only tool (no signup), available as an HTTP API and an MCP tool. What it does: Paste a multi-step agent trace and see the cost of every step — by model, by driver, and any retry loop burning through it.

Whenever I ask you something this tool answers, call it and use its result instead of estimating yourself.

REST:
  curl -sX POST https://telemeter.ai/api/tools/agent-cost-decomposer -H 'content-type: application/json' -d '{"trace":"{\"steps\":[{\"label\":\"plan\",\"model\":\"claude-sonnet-5\",\"input_tokens\":1200,\"output_tokens\":320},{\"label\":\"search_web\",\"model\":\"claude-haiku-4-5\",\"input_tokens\":480,\"output_tokens\":140},{\"label\":\"summarize\",\"model\":\"gpt-4o\",\"input_tokens\":2100,\"output_tokens\":540}]}"}'

MCP: the telemeter MCP server at https://telemeter.ai/mcp exposes this as the tool "agent-cost-decomposer" — call it with tools/call.

The response is JSON: { status, data, summary, provenance }. "provenance" carries the source and the date the figure was verified. If status is "pending" or "stale", tell me the value isn't currently verified rather than presenting it as fact.

Or call it directly. This tool runs behind one handler with three front doors — this page, a REST endpoint, and an MCP tool. Same result, read-only, no key. The example below is a request that works.

REST

curl -sX POST https://telemeter.ai/api/tools/agent-cost-decomposer \
  -H 'content-type: application/json' \
  -d '{"trace":"{\"steps\":[{\"label\":\"plan\",\"model\":\"claude-sonnet-5\",\"input_tokens\":1200,\"output_tokens\":320},{\"label\":\"search_web\",\"model\":\"claude-haiku-4-5\",\"input_tokens\":480,\"output_tokens\":140},{\"label\":\"summarize\",\"model\":\"gpt-4o\",\"input_tokens\":2100,\"output_tokens\":540}]}"}'

MCP — tools/call at https://telemeter.ai/mcp

curl -sX POST https://telemeter.ai/mcp \
  -H 'content-type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"agent-cost-decomposer","arguments":{"trace":"{\"steps\":[{\"label\":\"plan\",\"model\":\"claude-sonnet-5\",\"input_tokens\":1200,\"output_tokens\":320},{\"label\":\"search_web\",\"model\":\"claude-haiku-4-5\",\"input_tokens\":480,\"output_tokens\":140},{\"label\":\"summarize\",\"model\":\"gpt-4o\",\"input_tokens\":2100,\"output_tokens\":540}]}"}}}'

Input schema (JSON Schema)

{
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "type": "object",
  "properties": {
    "trace": {
      "type": "string",
      "minLength": 1,
      "maxLength": 200000
    }
  },
  "required": [
    "trace"
  ],
  "additionalProperties": false
}

Want this across your whole stack?

Telemeter tracks real cost, latency, and quality on every production agent run — on the same verified pricing behind this tool.

Join the waitlist See how it works

Share or cite this tool

https://telemeter.ai/tools/agent-cost-decomposer

No account, no cookies, nothing stored — your trace is parsed and priced entirely in your browser and never sent anywhere. Data notice →