GEPA

Free plan APISelf-hosted

What GEPA is

GEPA is an open-source Python framework for improving prompts and other text-based parts of software systems against evaluation metrics. It can optimize code, configurations, agent architectures, policies, and similar artifacts. Its optimize_anything API connects to a system through an evaluator you write, so use does not depend on DSPy or another framework. GEPA examines execution traces and evaluator feedback—including error messages, profiling data, or reasoning logs—to identify problems, suggest focused changes, and choose candidates with different strengths using a Pareto frontier. It can work with API-only models such as GPT, Claude, and Gemini without access to model weights.

The project provides adapters for workflows including DSPy, RAG, MCP, TerminalBench, and LangChain, and lists integrations with MLflow, Comet ML Opik, Pydantic AI, OpenAI Cookbook, Hugging Face Cookbook, and Google ADK. Human-readable traces explain why prompts changed and can help investigate agent behavior. Users can limit metric calls and set timeout, score, signal, file, composite, or no-improvement stopping conditions. The project says improvements may be possible with three examples and recommends 30–300 for best results. GEPA is free, self-hosted, and MIT-licensed; install it with pip.

Who it suits

GEPA suits developers and researchers optimizing systems where rollouts are expensive, examples are scarce, or models are available only through APIs. It is also a fit when interpretable traces are useful for understanding changes to prompts or agent behavior.

What it does well

Works through a user-written evaluator without requiring a specific framework.
Uses execution traces and diagnostic feedback to guide changes.
Supports API-only models without access to model weights.
Human-readable traces help explain prompt changes.
Offers metric-call caps and configurable stopping conditions.

What to know first

It does not inherently optimize for short prompts.
The project recommends 30–300 examples for best results.
Gradient-based methods remain effective with abundant data and 100,000 or more cheap rollouts.

Limits

The package is free under the MIT license. GEPA does not inherently optimize prompt length, and the project recommends 30–300 examples for best results; it also describes 100–500 evaluations as a use case for expensive rollouts.

Plans

Plan	Billed and limits	Price
GEPA	Open-source Python package · install with pip · MIT license	Free

As gepa-ai.github.io lists them, read 9 Oct 2026. The maker's own page has the final word.

Getting started

Visit the GEPA project website.
Install the Python package with pip install gepa.
Connect a system using optimize_anything and a user-written evaluator, or use a documented adapter.
Set metric-call limits and any stopping conditions for the optimization run.

Verdict

Choose GEPA if you want an open-source, self-hosted way to optimize text-based system components using evaluator feedback and readable traces. It may be a poor fit when shorter prompts are the priority unless prompt length is included as an objective, or when abundant cheap rollouts favor gradient-based methods.

Questions

How much does GEPA cost?

The listed GEPA plan is free. It is an open-source Python package under the MIT license.

Can GEPA work without DSPy?

Yes. Its optimize_anything API can work with a system through an evaluator you write, without requiring DSPy or another framework.

Which models can GEPA optimize?

It can optimize API-only models, including GPT, Claude, and Gemini, without access to model weights.

How many examples does GEPA need?

The project says it can show improvements with as few as three examples and recommends 30–300 for best results.

Where can I get help with GEPA?

The project directs users to Discord, Slack, GitHub Issues, and the project lead’s X account.

Facts from GEPA's own pages

Agent skill: The repository includes a GEPA Agent Skill for coding agents including Claude Code, Cursor, VS Code/Copilot, Codex, and Gemini CLI. · github.com, 9 Oct 2026
Best-fit workloads: GEPA is presented as a fit for expensive rollouts, scarce data, API-only models, or cases where human-readable optimization traces are useful. · gepa-ai.github.io, 9 Oct 2026
Budget controls: Users can cap metric calls and configure timeout, no-improvement, score-threshold, signal, file, and composite stop conditions. · gepa-ai.github.io, 4 Oct 2026
Built-in adapters: Built-in adapters include Default, Confidence, DSPy Full Program, Generic RAG, MCP, TerminalBench, AnyMaths, and LangChain adapters. · github.com, 4 Oct 2026
Comparison caveat: The site says gradient-based methods remain effective when data is abundant and there are 100,000 or more cheap rollouts. · gepa-ai.github.io, 9 Oct 2026
Data needs: The FAQ says GEPA can show improvements with as few as three examples and recommends aiming for 30–300 examples for best results. · gepa-ai.github.io, 4 Oct 2026
Data scale limit: The site says GEPA can work with as few as three examples and describes 100–500 evaluations as a use case for expensive rollouts. · gepa-ai.github.io, 9 Oct 2026
Development: GEPA is developed at UC Berkeley Sky Computing Lab through a research collaboration involving UC Berkeley, MIT, Stanford, and Databricks. · gepa-ai.github.io, 9 Oct 2026
Framework flexibility: The optimize_anything API can work with any system through a user-written evaluator, without requiring DSPy or another framework. · gepa-ai.github.io, 4 Oct 2026
How it works: It uses LLM reflection on execution traces to diagnose failures, propose targeted changes, and select candidates using a Pareto frontier. · gepa-ai.github.io, 9 Oct 2026
Installation: The project’s documented install command is `pip install gepa`. · github.com, 9 Oct 2026
Integrations: The project lists integrations with DSPy, MLflow, Comet ML Opik, Pydantic AI, OpenAI Cookbook, Hugging Face Cookbook, and Google ADK. · github.com, 4 Oct 2026
Interpretability: GEPA provides human-readable optimization traces that show why prompts changed and can help debug agent behavior. · gepa-ai.github.io, 4 Oct 2026
License: The GitHub repository lists the project under the MIT license. · github.com, 9 Oct 2026
Maker and origin: GEPA is developed at UC Berkeley Sky Computing Lab through a research collaboration between UC Berkeley, MIT, Stanford, and Databricks. · gepa-ai.github.io, 4 Oct 2026
Model access: GEPA can optimize API-only models such as GPT, Claude, and Gemini without requiring access to model weights. · gepa-ai.github.io, 9 Oct 2026
Notable limitation: GEPA does not inherently optimize for short prompts; its prompts may be longer unless prompt length is included as an optimization objective. · gepa-ai.github.io, 4 Oct 2026
Optimization method: It uses execution traces and evaluator-provided diagnostic feedback to guide LLM reflection, targeted mutations, and Pareto-aware candidate selection. · gepa-ai.github.io, 4 Oct 2026
Purpose: GEPA is a framework for optimizing prompts, code, agent architectures, configurations, and other text parameters against evaluation metrics. · github.com, 9 Oct 2026
Security and compliance: The opened official pages provide no security or compliance certification claims. · gepa-ai.github.io, 4 Oct 2026
Support: The project points users to its documentation, GitHub issues, and a Discord community. · gepa-ai.github.io, 9 Oct 2026
Supported systems: The project documents adapters for DSPy, RAG, MCP, TerminalBench, and LangChain. · gepa-ai.github.io, 9 Oct 2026
Trace feedback: Evaluators can provide error messages, profiling data, and reasoning logs as feedback for optimization. · gepa-ai.github.io, 9 Oct 2026
Use cases: The site lists production, coding agent, healthcare, research, and AI safety applications among its GEPA case studies. · gepa-ai.github.io, 9 Oct 2026

Where GEPA stands on Freedom251

AI Prompt Generators#6 of 28