claude-code-api-cost-estimator

Estimate your historical usage cost on third-party API providers (DeepSeek, Qwen, Kimi, or GLM)

Scans your local Claude Code sessions, buckets token usage to each provider's tiers, and calculates side-by-side cost comparison.

Quick start

npx claude-code-api-cost-estimator

The CLI starts a small local web server (default port 3000) and opens your browser to a comparison matrix.

If the browser doesn't open automatically (or you used --no-open), visit:

http://localhost:3000

Use a different port with --port:

npx claude-code-api-cost-estimator --port 4123
# then open http://localhost:4123

Stop the server with Ctrl+C in the terminal where it's running.

Requires Node.js 18 or newer.

If you'd rather clone and run:

git clone https://github.com/chrisjunlee/claude-code-api-cost-estimator
cd claude-code-api-cost-estimator
node bin/cli.mjs

Flags

Flag	Default	Description
--port <n>	3000	Port to listen on
--days <n>	30	Days of history to scan
--projects-dir <path>	~/.claude/projects	Where to find your .jsonl session files
--no-open	(off)	Don't auto-open the browser
-h, --help		Show help

Examples:

npx claude-code-api-cost-estimator --days 7
npx claude-code-api-cost-estimator --port 4123 --no-open
npx claude-code-api-cost-estimator --projects-dir /path/to/sessions

What it does

Claude Code stores every session as a JSONL file under ~/.claude/projects/. Each assistant line includes the model used and a usage block with input_tokens, output_tokens, cache_read_input_tokens, and cache-creation token counts.

This tool:

1. Walks ~/.claude/projects/**/*.jsonl (last N days)
2. Aggregates per-model token usage
3. Multiplies tokens by prices from pricing.json
4. Renders a side-by-side comparison: native Anthropic API cost vs. each third-party provider

The Anthropic API column is the baseline, not the point. The whole tool exists to put third-party numbers next to it so you can see the real delta on your workload, not a marketing chart.

Privacy

• All scanning and computation runs on your machine
• No transcripts, prompts, or token counts are sent anywhere
• The local server only binds to localhost

Updating pricing

All prices live in pricing.json at the package root. Edit it and refresh the browser. The server re-reads the file on every request, no restart needed.

Per-model fields (models):

• input / output: USD per 1M tokens
• cache_read: USD per 1M cache-read tokens (falls back to input if omitted)
• cache_write_5m / cache_write_1h: USD per 1M cache-write tokens (fall back to input)
• tiers: array of { max_input_tokens, input, output } for banded pricing (e.g. Qwen3-Max)
• tier: "premium" / "standard" / "fast" on Claude models; used for provider tier mapping
• valid_until: ISO date; shows a banner on the page while active

Provider entries (providers) let one dropdown option map Claude tiers to different underlying SKUs:

"providers": {
  "deepseek-v4": {
    "displayName": "DeepSeek V4",
    "tierMap": { "premium": "deepseek-v4-pro", "standard": "deepseek-v4-flash", "fast": "deepseek-v4-flash" }
  }
}

altOrder can reference either a model key (flat single-SKU) or a provider key (tier-mapped).

Contributing

Pricing changes and new providers are very welcome. Open a PR editing pricing.json with a link to the provider's official pricing page.

Provider pricing references currently used:

Provider	Pricing page
Anthropic	https://www.anthropic.com/pricing
DeepSeek	https://api-docs.deepseek.com/quick_start/pricing
Alibaba	https://www.alibabacloud.com/help/en/model-studio/models
Moonshot	https://platform.kimi.ai/docs/pricing/chat-k2
Zhipu Z.AI	https://docs.z.ai/guides/overview/pricing

Caveats

• Token counts come straight from Claude Code transcripts. Tokenizers differ across providers, so equivalent counts aren't perfectly equivalent.
• Cache-hit rates will vary in practice; the comparison assumes the same cache behaviour across all providers.

License

MIT