AI coding agents FAQ, answered with real run data
Short, sourced answers to the questions people ask about Claude Code, Codex CLI and Antigravity CLI. Every number comes from a logged 1P1A run: one frozen brief, one autonomous run, zero human help.
What is the best AI coding agent in 2026?
There is no reliable winner yet. In our 10 runs, Codex CLI shipped 3 of 3 apps, Antigravity CLI 3 of 3 and Claude Code 2 of 3. Antigravity CLI was fastest in every task. Codex CLI was the only agent to follow the expense-total spec exactly. Claude Code tested its own game most thoroughly, but also produced the only build that froze. We need at least 3 runs per stack before calling a winner. Full comparison.
Claude Code vs Codex CLI: which is better?
On the same 3 frozen specs, Codex CLI shipped all 3 and kept the expense total correct. Claude Code shipped 2 of 3: its kart racing game froze about 10 seconds into a race, and its expense tracker turned the grand total into a filtered subtotal. Claude Code self-tested more and ran longer (951.9s vs 208.5s on tower defense). Codex CLI produced the smaller game files. One run each.
Claude Code vs Antigravity CLI: which is better?
Antigravity CLI was faster in all 3 tasks: 200.0s vs 951.9s (tower defense), 102.2s vs 578.0s (kart racing) and 106s vs 171s (expense tracker). It shipped all 3; Claude Code shipped 2. Both made the same expense-total mistake, and Antigravity CLI also parsed saved data without a try/catch.
Codex CLI vs Antigravity CLI: which is better?
Both shipped all 3 tasks. Antigravity CLI was faster every time (200.0s vs 208.5s, 102.2s vs 522.0s, 106s vs 231s). Codex CLI was the only one that kept the expense total correct under a filter, and produced the smaller game files.
Which AI coding agent is the fastest?
Antigravity CLI with Gemini models, in all 3 tasks: 102.2 seconds for a Three.js kart racing game, 106 seconds for an expense tracker and 200.0 seconds for a tower defense game. Faster did not mean more correct: it shared the expense-total bug with Claude Code.
How much does Claude Code cost to build one app?
Claude Code reported $0.86 for a single-file expense tracker (claude-opus-5, 171s) and $0.74 for a single-file Three.js kart racing game (claude-sonnet-5, 578s). These are list prices from Claude Code's own log, not a subscription bill. Codex CLI and Antigravity CLI do not report a dollar cost on subscription billing, so we publish theirs as unknown instead of guessing.
Can an AI coding agent build a complete app from one prompt?
For small single-file apps, mostly yes: 9 of 10 runs shipped a working app from one complete brief with zero follow-up. But working is not the same as correct. In a 7-requirement spec, 2 of 3 agents silently changed what the total meant, and one game froze mid-play.
Do AI coding agents follow the spec exactly?
Not always. Our expense tracker spec asked for a running total of all expenses. Claude Code and Antigravity CLI computed it from the filtered rows, so filtering to Food showed $91.90 instead of the $1,365.90 grand total. Only Codex CLI kept it right. All three said they were done.
Does Claude Code test its own code?
Sometimes, and it shows. In the tower defense run it drove headless Chrome, played its own game, found a stalled wave counter and a stuck Start Wave button, and fixed both. That took 951.9 seconds, almost 5 times the fastest agent. In the kart racing run it did not catch an infinite loop that froze the game.
What kinds of bugs do AI coding agents leave behind?
- An infinite loop that froze a game about 10 seconds into play (Claude Code).
- A running total that quietly became a filtered subtotal (Claude Code, Antigravity CLI).
- Unguarded
JSON.parseof saved data that would crash the page if localStorage were corrupted (Antigravity CLI).
None of the 10 builds showed a console error on load. These bugs only appear when you use the app.
How do you benchmark AI coding agents fairly?
Same frozen brief for every agent, one autonomous run with tool approval on, zero human follow-up, and the full stack recorded: model, settings, harness version, permissions, environment, task version and run number. Unknown values are published as unknown. Reliability claims need at least 3 runs per stack and an automated test suite. Methodology.
Last updated Oct 31, 2026. Data: runs.json · llms-full.txt