Which AI Codes Best? Gemini 3 Pro, Opus 4.5 and Composer 1 Tested!

Quick Overview

Cursor with Composer 1 was the fastest overall, completing the complex iOS PDF scanner app task in 3.5 minutes and 1367 lines of code with only 4 prompts, significantly outperforming Gemini 3 Pro (9 minutes, 915 lines, 2 prompts) and Claude Opus 4.5 (12 minutes, 1685 lines, 1 prompt), though Gemini 3 Pro required fewer prompts than Cursor.

Key Points: Cursor using Composer 1 was the fastest, finishing the project in 3.5 minutes, generating 1367 lines of code using 4 total prompts. Gemini 3 Pro took 9 minutes to complete the task, generating 915 lines of code across 2 total prompts. Claude Opus 4.5 took the longest at 12 minutes, generating 1685 lines of code with only 1 prompt. The testing involved building a PDF Scanner iOS App based on detailed functional and design requirements documented in Markdown files. Cursor required multiple back-and-forth iterations (4 prompts) to fix an initial build error related to camera permissions. The experiment setup involved using Zapier MCP (Multi-Agent Coding Platform) to connect various AI coding agents (Cursor, Antigravity/Gemini 3 Pro, Claude Code/Opus 4.5) to a GitHub repository. The Google AI Pro plan costs $20/month and includes Gemini access, while Claude Pro costs $17/month (billed annually) for access to Opus 4.5.

Context: The video compares the performance of three different AI coding agents—Cursor (using Composer 1), Antigravity (using Gemini 3 Pro), and Claude Code (using Opus 4.5)—in tackling a complex software development task: building a feature-complete PDF Scanner iOS application from detailed requirements and mockups. The comparison tracks execution time, lines of code generated, and the number of prompts required to achieve the final working application, all orchestrated using Zapier's Multi-Agent Coding Platform (MCP).

Detailed Analysis

The video details a benchmark test comparing three AI coding agents—Cursor (Composer 1), Antigravity (Gemini 3 Pro), and Claude Code (Opus 4.5)—on the task of implementing a feature-complete PDF Scanner iOS application. The setup involved using Zapier's MCP to connect the agents to a GitHub repository containing detailed functional requirements, design mockups, and build steps. Cursor, leveraging Composer 1, was the clear winner in speed, completing the task in 3.5 minutes, generating 1367 lines of code across 4 prompts. Gemini 3 Pro required 9 minutes and 2 prompts, producing 915 lines. Claude Opus 4.5 took the longest at 12 minutes, generating 1685 lines, but used the fewest prompts (1). Cursor initially failed to build due to a camera permission error, requiring a few back-and-forth prompts to fix before the build succeeded. The final output for all models adhered to the functional and design requirements, but Cursor demonstrated superior efficiency in this specific coding scenario. The presenter also briefly reviewed the pricing structures for Google AI (Pro at $20/month) and Claude (Pro at $17/month annually).