PyTokenCalc

LLM token counting and cost estimation across cloud, local, and custom providers.

What it is

PyTokenCalc counts tokens across OpenAI, Anthropic, Google, Cohere, Azure-hosted OpenAI models, HuggingFace/open-source models, Ollama, and custom endpoints, then estimates the dollar cost of a request from a maintained per-model pricing table. It is pure Python — the speed comes from the tiktoken and HuggingFace tokenizers libraries it wraps, not from compiled code in PyTokenCalc itself. It installs and imports as pytokencalc.

Features / Highlights

• Multi-provider — OpenAI, Anthropic Claude, Google Gemini, Cohere, Azure-hosted OpenAI models, HuggingFace/open-source models, Ollama, and any custom HTTP endpoint you register.
• Provider-accurate tokenization — tiktoken for OpenAI and Azure-hosted OpenAI models, the HuggingFace transformers tokenizer, and live count-tokens endpoints for Anthropic/Google/Cohere.
• Pricing freshness — pricing_status() reports the snapshot's age and flags it stale after 90 days; load_pricing_override(path) loads a local JSON pricing file for newer/private models. pytokencalc pricing exposes it.
• Real cost estimation — estimate_cost() is backed by a per-model USD pricing table (input vs. output rates), not a guess; see pytokencalc/pricing.py for sources and last-updated date.
• Fast local counting — sub-millisecond for local (tiktoken/HuggingFace) providers on typical hardware.
• Batch processing — count tokens for many prompts in one call.
• Custom / BYOM models — register your own provider or fine-tuned model (see CUSTOM_PROVIDERS.md).
• Streaming/incremental counting — counter.streaming(model) gives a running token count as chunks of a live LLM response arrive, correct across chunk/BPE-merge boundaries (not a naive per-chunk sum).
• Encoding drift detection — get_tokenizer_info()["drift_warnings"] flags if an installed encoding's tokenization has changed since it was last verified.
• CLI — pytokencalc and the pycount shortcut, installed automatically.

Who it's for / Use cases

• Budget tracking — count a prompt and price the input/output before sending a request.
• Preventing cost overruns — reject requests above a cost threshold.
• Comparing providers — price the same prompt across models before choosing one.
• Streaming cost awareness — track tokens as a response streams in.

What it is not

• MCP transport is optional. pytokencalc.TokenCalculator has real handlers (token counts from the tokenizer registry, costs from the pricing table), and the FastMCP transport ships only with pip install "pytokencalc[mcp]"; without that extra, building a server raises an actionable install error. Budget/usage tools are in-memory and process-scoped.
• No committed benchmark result. The performance numbers are informal local observations, not a checked-in benchmark; there is no results file backing a specific claim. For OpenAI models PyTokenCalc is tiktoken under the hood, with a roughly 2–5% wrapper overhead, not a speedup.
• Lint/type coverage is partial. ruff check runs in CI, but black formatting and mypy typing are configured yet not enforced (separate passes).
• Anthropic/Google/Cohere counting makes a live network call to the provider's API, so those paths need the relevant package installed and an API key set.
• Provider coverage is a maintained table, not exhaustive. See docs/MODELS.md for exactly which models are covered.

Project status

Beta. Current version 1.4.1 (Development Status :: 5 - Production/Stable in pyproject.toml). The local OpenAI/HuggingFace counting path is well covered by tests; provider-specific paths require their optional dependencies. See ROADMAP_HONEST.md for the full honest status and technical-debt list.

MCP server

The 12 MCP tools are now served over a real FastMCP transport:

pip install "pytokencalc[mcp]"

# stdio (Claude Desktop, Cursor, etc.)
pytokencalc-mcp --transport stdio

# HTTP
pytokencalc-mcp --transport http --host 127.0.0.1 --port 8778

from pytokencalc import TokenCalculator

server = TokenCalculator().create_mcp_server()
server.run(transport="stdio")

The tools are the same real handlers documented in _mcp_tools.py; no external dab binary or separate service is involved. Budget/usage state is process-scoped, as documented.

Installation

pip install pytokencalc
# or with uv
uv pip install pytokencalc

# with local tokenizer support (tiktoken + HuggingFace transformers)
pip install "pytokencalc[tokenizers]"

Requires Python 3.9+.

Quick start

from pytokencalc import count_tokens, estimate_cost

# Count tokens instantly (runs locally via tiktoken — no API key needed)
tokens = count_tokens("Tell me a story about a robot", model="gpt-4o")
print(f"Tokens: {tokens}")

# Estimate cost from the pricing table
cost = estimate_cost("gpt-4o", input_tokens=tokens)
print(f"Cost: ${cost:.6f}")

Usage

Budget tracking (Anthropic/Google/Cohere models require the relevant package and an API key; OpenAI/GPT models work offline out of the box):

from pytokencalc import count_tokens, estimate_cost

prompt = "Hello"
reply_tokens_estimate = 50  # however you estimate expected output length

input_tokens = count_tokens(prompt, model="claude-3-opus")
cost = estimate_cost("claude-3-opus", input_tokens, reply_tokens_estimate)
print(f"Estimated cost: ${cost:.4f}")

Prevent overruns:

tokens = count_tokens(prompt, model="gpt-4")
if estimate_cost("gpt-4", tokens) > 0.10:
    print("Request too expensive, rejected")

Compare providers:

prompt = "Explain quantum computing in one paragraph."
for model in ["gpt-4o", "gpt-4-turbo", "gpt-3.5-turbo"]:
    tokens = count_tokens(prompt, model=model)
    cost = estimate_cost(model, tokens)
    print(f"{model}: {tokens} tokens, ${cost:.6f}")

Streaming/incremental counting — recomputes against the full accumulated text on each chunk, so it stays correct across BPE chunk boundaries:

from pytokencalc import TokenCounterRegistry

registry = TokenCounterRegistry()
stream = registry.get_counter("openai").streaming(model="gpt-4o")

for chunk in live_response_chunks:  # e.g. OpenAI SSE deltas
    delta = stream.add_chunk(chunk or "")
    print(f"+{delta} tokens, {stream.total_tokens} total")

Historical baseline estimates (opt-in, not exact) — if you have saved samples for a provider/model/task, estimate_from_baseline() extrapolates total tokens from the historical output/input ratio. This is a planning estimate, not a tokenizer count, and count_tokens()/estimate_cost() never use it:

from pathlib import Path
from pytokencalc import OKFTokenBaselines, TokenBaseline, estimate_from_baseline

store = OKFTokenBaselines(baseline_dir=Path("token_baselines"))
store.save_baseline(TokenBaseline(
    provider="openai", model="gpt-4o", task_type="chat",
    avg_input_tokens=100, avg_output_tokens=50, samples=200,
    percentile_95_input=250, percentile_95_output=140,
))

estimate = estimate_from_baseline("openai", "gpt-4o", "chat", 1000)
print(estimate.total_tokens)  # 1500 -- baseline estimate, not exact
print(estimate.p95_output_tokens)           # stored historical p95 output
print(estimate.predicted_p95_output_tokens) # p95 scaled to the 1000-token input

BaselineEstimate also exposes p95_output_tokens (the stored historical 95th-percentile output) and predicted_p95_output_tokens (that p95 scaled by input_tokens / avg_input_tokens, matching the mean prediction). Both are planning estimates for tail behavior, not guarantees.

Save baselines from real (prompt, output_tokens) samples with the CLI. Prompts are tokenized with the exact counter for the given model; output tokens should come from your own request logs. Samples are never fabricated or padded.

# JSONL: one object per line
# {"prompt": "Explain BPE", "output_tokens": 120}
pytokencalc baseline from-jsonl samples.jsonl \
    --provider openai --model gpt-4o --task-type chat

# JSON: an array, or {"provider": ..., "model": ..., "task_type": ..., "samples": [...]}
pytokencalc baseline save --samples samples.json \
    --provider openai --model gpt-4o --task-type chat

pytokencalc baseline list --provider openai
pytokencalc baseline compare --provider openai --task-type chat

baseline compare reports averages plus p95_input_tokens, p95_output_tokens, and p95_total_tokens for each model. All commands accept --baseline-dir DIR (default: ./token_baselines).

Returns None when no baseline exists. For exact counts use count_tokens().

CLI:

pytokencalc --help
pycount "Tell me a story about a robot" --model gpt-4o

Documentation

• API Reference — top-level functions, the full registry API, CLI, and REST server
• Supported Models — provider list, offline vs. API-backed, pricing table coverage
• Custom Providers — register your own endpoint or BYOM
• CLI Quick Start — the pycount terminal command
• Architecture — component overview, request flow, and honest architectural gaps
• Examples — runnable code samples
• Honest status & roadmap — what's tested, what isn't, known bugs, and technical debt

Known issues and deferred work

Blocked / failed:

• Automated upstream price feed (deliberately out of scope).
• Complete current-generation model coverage.

Rejected / deferred:

• Bundling the MCP transport (FastMCP) as a mandatory dependency.

Full shipped/blocked/rejected detail: QUEUE_STATUS.md.

Contributing

Contributions are welcome. See CONTRIBUTING.md and CODE_OF_CONDUCT.md. Report security issues per SECURITY.md.

pip install -e ".[dev,tokenizers]"
pytest tests/ -v --cov=pytokencalc
ruff check .

Related projects

• PyStreamXL — stream large Excel workbooks with constant memory
• PyStreamPDF — token-efficient PDF extraction for RAG
• PyStreamMCP — context and token optimization for AI agents
• PyStreamAI — ML deployment toolkit (canary/A-B, cost tracking, ONNX)
• PyTokenCalc — LLM token counting and cost estimation
• Browse all projects

License

This project is licensed under the Apache License 2.0.