benchgecko

Python client for BenchGecko, the AI model data platform. Look up benchmarks, compare models head-to-head, explore provider pricing, and estimate inference costs with a clean Pythonic API and zero dependencies.

BenchGecko tracks 414 models across 55 providers and 40 benchmarks, giving developers and teams the data they need to pick the right model for every task.

Installation

pip install benchgecko

Quick Start

import benchgecko as bg

# Look up a model
model = bg.get_model("claude-3-5-sonnet")
print(model["name"])                  # Claude 3.5 Sonnet
print(model["provider"])              # Anthropic
print(model["benchmarks"]["mmlu"])    # 88.7

# Estimate cost for a single call
cost = bg.estimate_cost("gpt-4o", input_tokens=4000, output_tokens=1000)
print(f"Total: ${cost['total_cost']}")  # Total: $0.02

API Reference

get_model(slug)

Returns the full data dictionary for a model, including benchmark scores and pricing. Returns None if the slug is not found.

gpt4o = bg.get_model("gpt-4o")
# {"slug", "name", "provider", "context_window", "input_price_per_1m", "output_price_per_1m", "benchmarks"}

compare_models(slug_a, slug_b)

Side-by-side comparison across every tracked benchmark, plus a pricing cost ratio. Useful for building comparison tables or making procurement decisions programmatically.

cmp = bg.compare_models("gpt-4o", "claude-3-5-sonnet")

# Per-benchmark delta
print(cmp["benchmarks"]["humaneval"])
# {"a": 90.2, "b": 92.0, "delta": -1.8, "winner": "claude-3-5-sonnet"}

# Which model is cheaper overall?
print(cmp["pricing"]["cheaper_model"])  # gpt-4o
print(cmp["pricing"]["cost_ratio"])     # 0.694

get_pricing(provider)

Lists every model from a provider with input/output pricing per million tokens and context window size.

for m in bg.get_pricing("anthropic"):
    print(f"{m['name']}: ${m['input_price_per_1m']}/M in, ${m['output_price_per_1m']}/M out")
# Claude 3.5 Sonnet: $3.0/M in, $15.0/M out
# Claude 3 Haiku: $0.25/M in, $1.25/M out

list_benchmarks()

Returns metadata for all tracked benchmarks: name, full name, description, and scoring scale.

for b in bg.list_benchmarks():
    print(f"{b['name']} ({b['full_name']}): {b['description']}")

estimate_cost(model, input_tokens, output_tokens)

Calculates the USD cost for a single inference call broken down by input and output tokens.

estimate = bg.estimate_cost("deepseek-v3", input_tokens=10_000, output_tokens=2_000)
print(estimate)
# {"model": "deepseek-v3", "input_cost": 0.0027, "output_cost": 0.0022, "total_cost": 0.0049, ...}

list_models() / list_providers()

Convenience helpers that return sorted lists of all available model slugs and provider keys.

print(bg.list_models())
# ["claude-3-5-sonnet", "claude-3-haiku", "command-r-plus", "deepseek-v3", ...]

print(bg.list_providers())
# ["anthropic", "cohere", "deepseek", "google", "meta", "mistral", "openai"]

Data Coverage

The bundled snapshot covers the most-used models from OpenAI, Anthropic, Google, Meta, Mistral, DeepSeek, and Cohere. For the full catalogue of 414 models, 55 providers, and 40 benchmarks, visit benchgecko.ai.

Pricing data and benchmark scores are updated with each package release. For real-time pricing, check the pricing page.

Requirements

Python 3.8 or later. No external dependencies.

License

MIT