Open-Source AI Framework GEPA Slashes Optimization Costs, Boosts GPT-4.1 Mini Performance by 10 Points

GEPA, a new open-source AI framework, is revolutionizing AI optimization by boosting GPT-4.1 Mini performance by 10 points and cutting cloud costs by 40%, while requiring a fraction of the evaluations needed by traditional reinforcement learning methods — already deployed across Shopify, Databricks, and 50+ production environments.

• GEPA (Genetic-Pareto) is an open-source AI framework that optimizes any textual parameter — including prompts, code, agent architectures, and configurations — using LLM-based reflection and Pareto-efficient evolutionary search, requiring only 100–500 evaluations compared to 5,000–25,000+ for reinforcement learning methods.
• The framework delivers significant real-world performance gains, including boosting GPT-4.1 Mini from 46.6% to 56.6% on AIME 2025, improving ARC-AGI agent accuracy from 32% to 89%, and achieving 40.2% cost savings in cloud scheduling — with adoption across Shopify, Databricks, Dropbox, OpenAI, and 50+ production deployments.
• GEPA is now integrated into major AI frameworks including DSPy, MLflow, Pydantic AI, Comet ML Opik, Google ADK, and the Gemini Enterprise Agent Platform, and supports a growing library of built-in adapters for RAG, LangChain, MCP, and mathematical reasoning tasks.