llm-cost

Offline token counter and cost estimator for LLM Engineering.

llm-cost is a statically linked CLI tool written in Zig. It replicates OpenAI's tiktoken logic with memory safety and offline capability. Designed for integration into CI/CD pipelines and infrastructure scripts.

Features

Performance: ~10 MB/s throughput on single core.
Offline: Embedded pricing and vocabulary; no value is sent over the network.
Portable: Static binary distribution for Linux, macOS, and Windows.
Parity: Validated against tiktoken using edge-case corpora (Unicode, Whitespace).
Control: Enforce cost limits via pipe mode.

Installation

Binaries Stable releases available on GitHub Releases.

Source Requires Zig 0.14.0.

git clone https://github.com/Rul1an/llm-cost
cd llm-cost
zig build -Doptimize=ReleaseFast
cp zig-out/bin/llm-cost /usr/local/bin/

Usage

Count Tokens

# Direct input
llm-cost count --model gpt-4o --text "Hello world"

# Pipe from file
cat document.txt | llm-cost count --model gpt-4o

Estimate Cost

llm-cost estimate --model gpt-4o --input-tokens 5000 --output-tokens 200

Analyze Corpus (Compression & Costs)

llm-cost report --model gpt-4o --json my_corpus.txt
# Output: {"stats":{...}, "metrics":{"bytes_per_token":4.2, "tokens_per_word":1.3}}

Pipeline Integration

# Fail if cost exceeds $1.00
cat logs.jsonl | llm-cost pipe --model gpt-4o --max-cost 1.00

Documentation

Project documentation follows the Diátaxis structure.

Type	Content
Guides	CI Integration, Release Verification
Reference	CLI Commands, Benchmarks, Man Page
Explanation	Architecture, Security Policy

Performance

Metric	Result (Apple Silicon)
Throughput	~10.11 MB/s
Latency (P99)	~0.13 ms (Small Inputs)
Complexity	O(N) Linear

See docs/reference/benchmarks.md for methodology.

Security

Builds adhere to SLSA Level 2 standards.

Secure Boot: Pricing database verified at runtime via Ed25519 Minisign signatures.
Artifacts: Signed with Cosign (Keyless via OIDC).
SBOM: CycloneDX format provided.
Reproducibility: Deterministic builds.

See docs/guides/verification.md for verification steps.

README