SWE-Bench Pro Accuracy vs Output Tokens Chart for GPT-5.x Models

Description

A line chart on a black background with the OpenAI logo in the top right corner, titled 'SWE-Bench Pro (Public)'. The Y-axis shows Accuracy from 30% to 60%, and the X-axis shows Output tokens from 0 to 100,000. Three model curves are displayed: GPT-5.3-Codex (white/lightest line) achieves the highest accuracy around 57% with roughly 25,000-30,000 output tokens before plateauing, GPT-5.2-Codex (medium blue) reaches about 55% accuracy but requires around 90,000 tokens, and GPT-5.2 (darker blue) tracks slightly below GPT-5.2-Codex. The chart demonstrates that GPT-5.3-Codex achieves better results with significantly fewer tokens, showing improved efficiency