budget-forge

CLI tool that finds the optimal thinking budget for reasoning LLMs — test prompts at multiple budget levels and generate cost-vs-quality tradeoff curves

Budget Forge

Find the optimal thinking budget for reasoning LLMs — maximize quality per dollar spent.

Budget Forge is a CLI benchmarking tool that sweeps across thinking budgets for reasoning models (DeepSeek R1, Qwen3, Claude Sonnet, o3-mini), measures quality and cost at each level, and identifies the sweet spot where you get the most quality per penny.

Why?

Reasoning models can think before they answer — allocating internal “thinking tokens” to reason through a problem. This improves quality, but also costs more. The relationship between thinking budget and response quality is not linear:

  • Too few thinking tokens → the model rushes, quality suffers
  • Too many thinking tokens → you burn money on overthinking
  • Somewhere in between lies the sweet spot — the budget that gives you the best quality-to-cost ratio

Budget Forge makes finding that sweet spot systematic, reproducible, and automated.

Installation

pip install budget-forge

Or from source:

git clone https://github.com/vikasudasi/budget-forge.git
cd budget-forge
pip install -e .

Usage

Basic benchmark

budget-forge test \
  --prompt "Explain memoization with an example." \
  --model deepseek-r1

Custom budgets

budget-forge test \
  --prompt "Write a Python function for merge sort." \
  --model qwen3 \
  --budgets 0,512,1024,2048,4096,8192

CSV export

budget-forge test \
  --prompt "Explain the HTTP protocol." \
  --model deepseek-r1 \
  --csv results.csv

Prompt file & reference answer

budget-forge test --prompt-file prompt.txt --model claude-sonnet
budget-forge test --prompt "What is 2+2?" --reference "4" --model o3-mini

Supported Models

Model CLI flag API base Env var
DeepSeek R1 deepseek-r1 api.deepseek.com DEEPSEEK_API_KEY
Qwen3 qwen3 api.openrouter.ai OPENROUTER_API_KEY
Claude Sonnet claude-sonnet api.anthropic.com ANTHROPIC_API_KEY
o3-mini o3-mini api.openai.com OPENAI_API_KEY

How It Works

For each budget level, budget-forge sends your prompt with the thinking budget parameter specific to each model’s API. It records thinking tokens, total tokens, cost, and latency. Quality is scored by self-evaluation (1-10) or exact match against a reference answer. The sweet spot is the budget where quality per dollar peaks.

License

MIT