Video summary

I Made Codex and Claude Code Build the Same App. One Clearly Won.

Main summary

Key takeaways

Product Review

Product Reviewed

Two AI coding agents generated the same production-ready “Typeform alternative” form builder app:

  • Codex (“Real Form” output)
  • Claude Code (“Formora” output)

Key Features of the Generated Apps (as seen in the demo)

Codex result: “Real Form”

  • Admin/workspace demo area with a form builder supporting multi-step capabilities (multi-page forms)
  • Ability to insert many form elements, including:
    • Text
    • Variables (e.g., rating/price)
    • Images
    • Matrices
    • Uploads
  • Advanced features supported:
    • Variables
    • Logic / conditional routing
    • Theming controls (e.g., corner radius and colors)
    • Webhooks and email notifications
    • Sharing links
    • Accessibility checking
    • Partial responses
  • Form UI included enhancements such as:
    • Welcome screen customization (e.g., image upload, alt text, focus/position)

Claude Code result: “Formora”

  • More user-friendly, Typeform-like experience
  • Supports conversational vs multi-layered display modes
  • Includes UI niceties such as:
    • Progress indicator
    • Question numbers
    • Keyboard hints
    • Auto-save
    • Partial response commits
  • Question types supported:
    • Short/long text
    • Email/phone/website
    • Dropdown
    • NPS
    • Opinion scale
    • Rating
    • Ranking
    • Image selection
  • Supports:
    • Mandatory questions
    • Character limits/patterns
  • Supports logic / branching (redirects based on answers)
  • Design themes supported (including dark theme) plus sharing/publishing workflow
  • Results viewing (summary/answers)

User Experience: What Worked Well vs. What Didn’t

Codex (Real Form)

Pros

  • Visually polished at first glance; main UI styling feels solid
  • Strong admin-side data modeling depth

Cons

  • Admin/user interface feels overwhelming/confusing
    • Too many controls at once
    • Difficulty deleting fields
  • Multiple UI bugs / quality issues were encountered:
    • Image preview not available after adding an image
    • Theme changes not responding (colors/corner radius appear not to work)
    • Confirmation popup appears offset/left
  • Demo functionality felt incomplete relative to what a fully finished product would be expected to deliver

Claude Code (Formora)

Pros

  • Much easier to understand and use (“I know exactly what to do”)
  • Closer to the Typeform interaction model (one-question-at-a-time, conversational feel)
  • Generated form + results flow is easier to navigate
  • Better overall functionality for the creator workflow compared to Codex’s first build

Cons

  • Still had some bugs:
    • Account/dashboard UI issue where a button behavior was unclear
    • Navigation friction after certain actions (e.g., can’t easily go back after design/webhooks)
    • Logic/UI counters became inconsistent later in testing (“one of three”, “one of two”, etc.)
    • Website field type didn’t render as expected until switching/retrying

Comparisons Made

  • The video is explicitly a side-by-side comparison of Codex vs Claude Code producing the same Typeform-clone app.
  • Comparison dimensions included:
    • Design / user experience
    • Functional completeness
    • Cost
    • Time
    • Agent/tool usage complexity
    • Testing coverage and reliability
    • Qualitative agent behavior (planning/creativity vs obedience/execution)

Performance + Cost + Effort (Numerical Results)

Runtime / Cost

  • Claude Code

    • Time: ~5.5 hours
    • Cost: $832 (with a note of possible discrepancy; later stated closer to ~$800)
    • Efficiency: described as ~11x faster and ~6.6x cheaper

  • Codex

    • Time: ~61 hours (~2.5 days)
    • Cost: ~$3,000

Tokens / Scale

  • Claude Code output tokens: just over 2 million
  • Codex output tokens: almost 11.5 million
  • Model/orchestration notes:
    • Codex used GPT-5.6-Sol (via orchestration)
    • Claude Code used a mix (including Fable and Opus, with Opus 4.8 becoming the main orchestrator after a security check)

Architecture / Tooling Complexity

  • Claude Code:
    • 1 orchestrator
    • 35 subagents
    • ~2,800 tool calls
  • Codex:
    • 1 orchestrator
    • 126 subagents
    • ~32.5k tool calls

Testing Performed (unit / unit+browser)

  • Claude Code:
    • 296 unit tests
    • 199 test cases
    • 102 browser tests
  • Codex:
    • 2300 unit tests
    • 341 test cases
    • 391 browser tests

Ratings / Scores (as stated)

  • “Efficiency”:
    • Claude Code: 9.8/10
    • Codex: 5.5/10

Main Takeaway / Verdict

  • Claude Code wins on output quality for this specific app: the final “Formora” experience is more Typeform-like, more usable, and judged to have better actual output.
  • Codex wins on testing and reliability infrastructure: it ran far more tests, including broader browser coverage and more aggressive testing strategies (cross-browser, property tests, bug injection mentioned).

Overall implied recommendation:

  • Choose Claude Code if you want a better user-facing app faster and cheaper.
  • Choose Codex if you want heavier testing and robustness/coverage, but expect higher cost/time and potentially worse initial UX.

Unique Points Mentioned (Consolidated List)

  1. Both agents were tasked to build the same production-ready Typeform clone using /target.
  2. Codex took ~3 days / 61 hours and cost ~$3,000; Claude Code took ~5.5 hours and cost ~$832 (≈ $800).
  3. Claude Code’s “Formora” had a clearer, Typeform-like conversational UI and workflow.
  4. Codex’s “Real Form” admin/user interface was confusing/overwhelming and had several UI bugs.
  5. Both included advanced features: multi-page forms, logic/branching, theming, sharing, webhooks, and more.
  6. Codex emphasized deeper internal capabilities (more subagents/tools/tokens) but produced a less polished UX.
  7. Claude Code used fewer agents/tools and produced better functional output in this run.
  8. Testing coverage favored Codex heavily (more unit/browser/property-style testing).
  9. Characterization of agent style: Claude Code felt more “creative/planning,” while Codex felt more “obedient/execution-focused.”
  10. The builder recommended continuing iteration beyond the first successful build (conceptual guidance used in the prompt).

Speaker-Specific Views (End)

  • A single narrator throughout, but with different “judges” expressed within the same voice:
    • User-experience perspective during demo usage
    • Quantitative analysis (cost/time/tokens/tools/tests)
    • Meta commentary on agent behavior (creativity/planning vs obedience/testing rigor)

Overall recommendation: For a Typeform-like app that you want working quickly with a better creator/UI experience, Claude Code is the stronger choice in this experiment. For maximum test coverage and reliability-oriented engineering, Codex is stronger but slower and costlier.

Original video