Running Antigravity 2.0 and Codex in Parallel for One Month: A Practical Selection Line for Indie Developers
A May 2026 head-to-head running Antigravity 2.0 and Codex on the same indie-development work, with the selection criteria and cost reversal line I now use.
In May 2026 I ran the same work through Antigravity 2.0 and OpenAI Codex side by side. There is no shortage of comparison posts about either tool, but I wanted my own ground truth on "when does each one win" from real indie development. The work I used was a StoreKit 2 migration across four iOS wallpaper apps on a single Mac mini M5, and the picture that emerged was less binary than I expected.
I am Masaki Hirokawa, an artist and indie developer running wallpaper apps since 2014 (over 50 million downloads across the portfolio) and four AI-tech blogs (Claude Lab, Gemini Lab, Antigravity Lab, Rork Lab) on autopilot. The story below is from the StoreKit 2 sweep on the wallpaper apps.
Why I ran them in parallel
In early May, I needed to grep SKPaymentQueue traces and apply the same diff pattern across four repos. That is the kind of work AI coding assistance helps with the most. I normally lean on Antigravity 2.0, but I had heard the recent Codex update changed the terminal-integration feel, so I committed to running the same prompts through both in parallel for a full week.
The rules were simple:
Send identical prompts to both tools (same wording).
If one stalls, hand the same task to the other.
Log success/failure and elapsed time for every task.
Three days and 47 tasks later, the data was enough to settle the selection question.
The selection line that emerged
The short version of where the line falls in my workload:
Four-repo cross-cutting greps and rewrites: Antigravity 2.0 wins.
Detailed in-file behavior fixes: Codex wins.
Investigations that lean on official documentation: Antigravity 2.0 wins.
First-draft test code generation: roughly a tie.
Interactive in-editor exploration: Codex wins.
In numbers: cross-repo tasks were about 25% faster on Antigravity 2.0; small in-editor conversational fixes were about 18% faster on Codex. Overall the split came out roughly 60 / 40 in Antigravity's favor.
✦
Thank you for reading this far.
Continue Reading
What follows includes implementation code, benchmarks, and practical content we hope you'll find useful. This site runs without ads — server and development costs are supported entirely by members like you. If it's been helpful, we'd be truly grateful for your support.
WHAT YOU'LL LEARN
✦Head-to-head numbers from running both tools on the same May 2026 workload
✦An indie-developer cost reversal line in real revenue terms
✦Behavioral differences observed across single-file vs four-repo parallel work
Secure payment via Stripe · Cancel anytime
✦
Unlock This Article
Get full access to the rest of this article. Buy once, read anytime. This site is ad-free — your support goes directly toward keeping it running.
Costs for the May parallel run came in at about 12,500 yen on Antigravity 2.0 and about 9,800 yen on Codex. Codex was cheaper in absolute terms, but that reflected the heavy share of small in-editor work; the cross-repo work that only Antigravity 2.0 finished cleanly is not in that comparison.
In indie economics, you start feeling AI coding spend once monthly total crosses around 20,000 yen, because that becomes a noticeable percentage of AdMob revenue. The wallpaper-app portfolio generates well over six figures yen per month from AdMob, so 20,000 yen is acceptable. If your monthly indie revenue is in the 30,000 yen range, picking one tool is more realistic.
When parallel operation stops paying off
Parallel running is genuinely fun for the first couple of days. You see real-time behavior differences and your mental model of each tool fills in fast. From day 3, though, the overhead starts to bite.
The reversal points I noticed:
Through day 2, parallel running maximizes learning speed.
From day 3, the cost of deciding "which tool gets this task" becomes a real tax.
By day 5, you have days where you carry both but only actually use one.
My personal landing was to put Antigravity 2.0 back as the default from day 4 onward, with Codex as the lightweight assistant. That choice preserves focus far better than a permanent dual-tool workflow.
Example: same prompt, different behavior
The most divergent prompt of the run was:
"List every occurrence of SKPaymentQueue.default() across the four repos, and for each location propose a migration to Transaction.updates on a per-file basis. Do not touch UI code or the initial launch sequence (@main)."
Antigravity 2.0: listed 17 occurrences, attached 5 lines of context to each, in 4 minutes.
Codex: listed 8 occurrences in one repo and stopped, asking for a separate prompt for the other three repos. Total 7 minutes including the follow-up prompt.
That gap reflects how cleanly Antigravity 2.0 handles multi-file, multi-repo assumptions. On a different prompt - "rewrite Entitlement.refresh so that both subscription and AdMob reward signals are always evaluated, not short-circuited" - Codex landed the in-file rewrite in one pass, while Antigravity inserted one extra confirmation step.
Where I am landing for now
My current setup keeps Antigravity 2.0 as the default and switches to Codex only for in-editor, immediate-feedback work. With the behavior differences from May in mind, the switch decision takes about 5 seconds.
For anyone about to try both, I would skip the "pick one" framing entirely. Run them in parallel for three days, log behavior differences, then collapse to a single default on day 4. Three days is enough to fingerprint each tool against your own workload; from day 4, the overhead of parallel runs outweighs the learning.
Tool selection in indie development is not a once-a-year choice. The right answer shifts with the work, and as a one-person operation supporting a 50-million-download wallpaper portfolio, having a stable selection rubric to fall back on keeps the noise down. I hope this is useful as you build yours.
Share
Thank You for Reading
Antigravity Lab is ad-free, supported entirely by members like you. We publish practical guides daily with implementation code, benchmarks, and production-ready patterns. If you've found it useful, we'd love to have you on board.