Claude Opus 5 vs GPT-5.6 Sol: Same App, Same Prompt

This video compares Anthropic’s Claude Opus 5 and OpenAI’s GPT-5.6 Soul by building the exact same React Native Expo app using the same prompt. Both models are tested on their ability to deliver a working app, developer experience, and cost efficiency.

I show how Opus 5 handles the build process, including using the iOS simulator for validation, and test the app on a physical device. The video also covers the cost and time spent, plus insights on marketing the app using AI-driven social data.

What's inside

• Introduction to Claude Opus 5 and GPT-5.6 Soul comparison
• Using bypass permissions for automated agent approval
• Pricing and cost comparison between Opus 5 and GPT-5.6 Soul
• Opus 5 running the React Native Expo app build and simulator testing
• Real device testing and app functionality review
• Limitations of simulator testing and manual testing needs
• Cost and time analysis of the build session
• Using Viralo data to create a launch plan for the app

Introduction to Claude Opus 5 and GPT-5.6 Soul comparison

Claude Opus 5 is Anthropic’s latest model positioned as a direct competitor to OpenAI’s GPT-5.6 Soul. Both models have similar input pricing, but Opus 5 offers cheaper output tokens. Anthropic claims Opus 5 is better at verifying code, debugging, and iterating until success, while GPT-5.6 Soul previously showed a better developer experience and app quality in tests.

In this video, I test both models on the same React Native Expo app prompt to see which can actually build a real app better, not just benchmark well. The app is a habit tracker named "Tally," and I use the same skills and MCP server integration for both.

Using bypass permissions for automated agent approval

To speed up the build process, I enable bypass permissions in Opus 5. This lets the AI agent handle all approvals automatically without manual intervention. It’s crucial for letting the agent run through the entire build, test, and validation workflow without stopping for permission prompts.

This setup simulates a real-world scenario where the AI can autonomously create and verify the app, making it easier to compare the models fairly and see how well Opus 5 manages the full development lifecycle.

Pricing and cost comparison between Opus 5 and GPT-5.6 Soul

Both Opus 5 and GPT-5.6 Soul cost about $5 per million input tokens. Opus 5 charges $25 per million output tokens, slightly cheaper than GPT-5.6 Soul’s $30. This makes Opus 5 roughly 50% cheaper than Fable 5, Anthropic’s previous model, while also outperforming it in coding tasks.

During the session, the total cost for Opus 5 was about $73 over an hour, mostly spent validating the app using the iOS simulator. For comparison, building the same app with Fable 5 cost me $50. The pricing shows Opus 5 is competitive but still expensive for long sessions.

Opus 5 running the React Native Expo app build and simulator testing

Opus 5 builds the React Native Expo app from a blank template, using my app icon skill and MCP server for package info and implementation details like haptics and widgets. It names the app "Tally," matching previous tests.

A new feature is Opus 5’s ability to run the iOS simulator inside Claude code, taking screenshots and analyzing the UI to detect layout issues. For example, it found a summary block overlapping the header and automatically rewrote the home screen to fix it.

This simulator integration is a game changer because the AI can see what the app looks like and validate UI correctness during the build, not just generate code blindly.

Real device testing and app functionality review

After the build, I test the app on a physical iOS device. The onboarding flow works, including dark mode support and notifications. The app’s navigation feels smooth, and sample data loads correctly.

Some minor issues remain: the title is not compatible with light mode, and the header animation does not work. However, sorting, marking habits done, and local notifications function well. The new habit screen is minimalistic and clean.

Swipe actions to edit or delete items also work perfectly. Overall, the app is functional and close to production-ready, demonstrating Opus 5’s strong coding and UI capabilities.

Limitations of simulator testing and manual testing needs

While the iOS simulator testing is impressive, it has limitations. Opus 5 cannot test widgets fully, activate test notifications, or unblock the simulator after push notification tests. Complex gestures like zoom, drag-to-reorder, and transitions require manual testing.

This means AI-assisted builds still need human QA for advanced UI interactions and edge cases. Simulator screenshots help catch layout bugs early but don’t replace thorough manual testing.

Cost and time analysis of the build session

The session took about an hour, with most time spent validating the app’s functionality using the simulator. Compiling the app took 2-3 minutes per iteration, adding to the total time.

The $73 cost is high for a single app build, especially compared to $50 spent on Fable 5 for the same app. However, the $20/month Claude Pro subscription covers the base usage, so the extra cost is from token consumption during the session.

The context window used about 55% of the 1 million token limit, showing the complexity of the task and the amount of information processed.

Using Viralo data to create a launch plan for the app

To help with app discovery, I partnered with Viralo, a tool that tracks creators and content trends across TikTok, YouTube, Instagram, and Meta Ad Library. Viralo connects to Claude via MCP, letting Opus 5 research real social data on habit and productivity apps.

Viralo found that iPhone home screen aesthetics and widgets have more viral content than dedicated habit tracker demos. Based on this, Opus 5 generated three launch concepts: a home screen carousel, a silent 8-second demo, and a simple builder story asking for feedback.

This data-driven approach helps create a launch plan grounded in actual social trends, not guesswork, improving chances of app success.

Resources

CourseReact Native course

Fundamentals through shipping: the concepts behind the prompts, with lifetime access.

Premium resourcePro membership

Access Claude Pro and other AI tools to build and test apps with advanced AI models.

InkigoInkigo: AI-powered app for content research

Connect Viralo and other AI tools to your projects for data-driven insights.