Claude Code vs Codex CLI vs Antigravity CLI: same prompt, real results
We gave three AI coding agents the same complete product brief for three different apps, one autonomous run each and zero human help. Here is what each one shipped, how long it took, what it cost and what it got wrong.
Ship rate across 3 apps
| Agent (harness) | Tower defense | 3D kart racing | Expense tracker | Shipped |
|---|---|---|---|---|
| Claude Code | Ship | No-ship (freezes) | Ship, total bug | 2 / 3 |
| Codex CLI | Ship | Ship | Ship, spec-exact | 3 / 3 |
| Antigravity CLI | Ship | Ship (2 models) | Ship, total bug + unsafe parse | 3 / 3 |
Antigravity CLI ran the kart task twice, with Gemini 3.7 Flash and Gemini 3.8 Flash. Both shipped. Models differed between tasks; see the full table below.
All 10 runs
| Task | Stack | Verdict | Wall time | Output tokens | Reported cost |
|---|---|---|---|---|---|
| Tower defense | Antigravity CLI + gemini-3.7-flash-medium | Ship | 200.0s | 65,049 | unknown |
| Tower defense | Codex CLI 0.151.0 + gpt-5.6-luna (medium) | Ship | 208.5s | 8,630 | unknown |
| Tower defense | Claude Code 2.1.252 + claude-sonnet-5 | Ship | 951.9s | unknown | unknown |
| 3D kart racing | Antigravity CLI + gemini-3.7-flash (medium) | Ship | 102.2s | 25,734 | unknown |
| 3D kart racing | Antigravity CLI + gemini-3.8-flash-medium | Ship | 292.3s | 87,460 | unknown |
| 3D kart racing | Codex CLI + gpt-5.6-luna (medium) | Ship | 522.0s | 26,036 | unknown |
| 3D kart racing | Claude Code + claude-sonnet-5 | No-ship | 578.0s | 62,317 | $0.74 |
| Expense tracker | Claude Code 2.1.252 + claude-opus-5 | Ship | 171s | 11,916 | $0.86 |
| Expense tracker | Codex CLI 0.151.0 + gpt-5.6-sol (medium) | Ship | 231s | 8,758 | unknown |
| Expense tracker | Antigravity CLI + gemini-3.1-pro-high | Ship | 106s | 12,076 | unknown |
Codex CLI and Antigravity CLI report tokens but no dollar cost on subscription billing, so cost is unknown, not estimated. The tower defense Claude Code log did not keep its usage block. Full logs and corrections are kept with each run.
Claude Code: the most thorough, and the only freeze
- Slowest in both game tasks (951.9s and 578.0s). In tower defense it drove headless Chrome, played its own game, found 2 bugs and fixed them before stopping.
- Its kart racing game freezes about 10 seconds into a race because of an infinite loop in the AI steering code. We first published a patched copy without saying so, and corrected that on Oct 31, 2026.
- In the expense tracker it turned "running total of all expenses" into a filtered subtotal.
- The only CLI of the three that reports a dollar cost: $0.74 and $0.86 on the runs where the log kept it.
Codex CLI: smallest game files, the only spec-exact total
- Shipped all 3 tasks. Produced the smallest file in both game tasks (20 KB tower defense, 25 KB kart racing), packed into very long lines.
- The only agent that kept the expense total as a total of all expenses under a filter, as the spec said.
- Mid-pack on speed: 208.5s, 522.0s and 231s.
Antigravity CLI: fastest in all 3 tasks
- Fastest run in every task: 200.0s, 102.2s and 106s.
- Shipped every task, with two different Gemini Flash models on kart racing.
- In the expense tracker it made the same total mistake as Claude Code, and it parsed saved data without a try/catch, so corrupted localStorage would crash the page on load.
What this does and does not tell you
- It does show real, reproducible differences in how each agent works: how much it self-tests, how long it runs and where it drifts from the spec.
- It does not show which agent is more reliable. That needs at least 3 runs per stack on the same task and an automated test suite. Why.
- It does not compare equal models. Each harness ran its own vendor's model, and model tiers differed by task.
What we test next
A real business app spec with an automated test suite and 3 runs per stack, plus a rerun on the September and October 2026 releases (Claude Opus 5.5 and Sonnet 5.5 in Claude Code, GPT-6.1 Sol in Codex CLI, Gemini 4 Argon in Antigravity CLI). Follow @tomcandev for results.
Data from runs on Sep 1 to 3, 2026. Corrections applied Oct 31, 2026. Raw data: runs.json · Quick answers: FAQ