I'm excited to see OpenAI, Anthropic, and Google pushing each other like this. Quick test of GPT-5.2 with the Codex plugin in VS Code on an Ubuntu device.

The Task

I had a static HTML page for tracking my daughter's 4th-grade basketball games and asked GPT-5.2 to connect it to Firebase. The page was part of AllPlays, a static HTML + JavaScript sports team management and stat tracking application I've been building.

The beta mock was an in-memory prototype with no backend. I needed it wired to the same Firebase patterns used by the existing track.html page: authentication, Firestore reads/writes, and the same data schema for events and aggregated stats.

The Experience

GPT-5.2 asked a few clarifying questions and, with minimal back and forth, it delivered a working integration. The workflow felt natural: it understood the existing codebase structure, asked about authentication patterns, and produced code that matched the conventions already in place.

Initial Concerns

I was initially concerned about performance. The Codex plugin in VS Code can sometimes feel sluggish on larger tasks, and I wasn't sure how GPT-5.2 would handle the context of an existing Firebase-backed application with multiple HTML pages and shared JavaScript modules.

What GPT-5.2 Built

The integration followed the same patterns as the existing tracker:

  • Auth + bootstrap: checkAuth() redirect to login, read teamId/gameId from URL hash, load team/game/players in parallel
  • Dynamic stat columns: Read from statTrackerConfigId in Firestore, fall back to default basketball columns (PTS, REB, AST, STL, TO)
  • In-memory tracking: All stats stay local until the user hits finish, avoiding noisy realtime writes
  • Batch write on finish: Events to teams/{teamId}/games/{gameId}/events, aggregated stats per player, game document update with scores and status
  • AI summary + email: Lazy-load Firebase AI for Gemini-powered match summaries, generate mailto draft for email recaps

The Safety Pattern

One detail I appreciated: GPT-5.2 replicated the existing "clear prior data" safety check. On first Start, if Firestore already has events/stats from a previous session, it prompts the user to clear before overwriting. Small thing, but it shows the model was paying attention to the existing code conventions.

The basketball tracker is now live at allplays.ai, and you can see the code and documentation in the GitHub repo.

Competition between OpenAI, Anthropic, and Google is producing rapid improvements. Each release pushes the others to respond, and we're all benefiting from it.

Takeaway

This was a quick test, not a deep evaluation. But for a task like porting a beta mock to production Firebase with existing conventions to follow, GPT-5.2 with Codex handled it well. Worth keeping an eye on as the model matures.

GPT-5.2 OpenAI Codex Firebase VS Code