How We Cut Claude Code Token Usage by 94%

A reproducible benchmark on FastAPI showing exactly how Code Context Engine reduces input token costs for AI coding agents.

The problem: input tokens are 85-95% of your bill

If you use Claude Code, Cursor, or any AI coding agent, most of your spend goes to input tokens. Every time Claude needs context about your codebase, it reads entire files. A question about one function pulls in 800 lines.

We built Code Context Engine (CCE), a local MCP server that indexes your codebase and returns only the relevant chunks. No cloud. No API key. Your code stays on your machine.

But "it saves tokens" isn't a benchmark. Here's how we actually measured it.

Benchmark setup

We tested against FastAPI (53 source files, 180,000 tokens total) with 20 real coding questions. Not synthetic. Not cherry-picked.

What "without CCE" means

For each query, we measure the total tokens if Claude reads the full content of every file the query touches. This is the baseline.

Average without CCE: 83,681 tokens per query.

This is conservative. In practice, agents often read more files than strictly needed (exploring, scanning imports, checking related code).

What "with CCE" means

Claude calls context_search("payment flow"). CCE returns the relevant code chunks with confidence scores. Only the code that answers the question.

Average with CCE: 4,927 tokens per query.

That's a 94% reduction in input tokens.

Per-layer savings

CCE has multiple layers, each measured independently against its own baseline:

These don't stack multiplicatively. The headline number is 94% from retrieval alone.

Retrieval quality: Recall@10

Saving tokens is useless if the wrong code is retrieved. We measured Recall@10: how often the correct file appeared in the top 10 results.

Result: 0.90. Nine out of ten times, CCE found the right file. The 10% miss rate is usually edge cases where the answer spans 4+ files.

What this means for your Claude Code bill

On Sonnet pricing ($3/1M input tokens), a medium project session without CCE costs roughly $0.14. With CCE: $0.04. That's a 70% cost reduction on the session.

On Opus ($15/1M), the savings are more dramatic: $0.70 per session down to $0.21.

Limitations (honest)

Single repo. This is FastAPI only. Different codebases (monorepos, many small files, heavily templated code) may see different numbers.

Python only. CCE supports JS, TS, Go, Rust, Java, and PHP for chunking, but this benchmark only tested Python.

Open-source embedding model. Cloud services using larger models likely have better retrieval on ambiguous queries.

Reproduce it yourself

pip install code-context-engine
python benchmarks/run_benchmark.py \
  --repo https://github.com/fastapi/fastapi.git \
  --source-dir fastapi

Takes about 60 seconds. Outputs a markdown file with per-query token counts, recall scores, and latency measurements.