Automatic Prompt Optimization

# Automatic Prompt Optimization — 2026

> **The best prompt writer is now a machine.** Models with evolutionary loops and eval sets rewrite instructions better than expert humans — by a wide margin on hard tasks. The skill is shifting from writing prompts to defining the objectives the optimizer writes against.

## The Shift

In 2026, automatic prompt optimization (APO) has moved from research to production:

- **GEPA** (ICLR 2026): Evolutionary optimizer lifted a CoT program on MATH from 67% → 93% through instruction refinement alone — no few-shot examples, no fine-tuning
- **Shopify**: Moved a GPT-5 prompt to a small open model optimized with GEPA — 75x cheaper, 2x more reliable
- **DSPy ecosystem**: Mature tooling for programmatic prompt optimization

---

## How APO Works