I asked Claude Code, Codex, and Antigravity to build the same game, and one absolutely crushed the others

Two weeks ago, a number was published stating that a Claude Code subagent costs around 436,000 tokens before it reads a single file. The same measurement method was repeated in two articles, but today a cleaner measurement yielded a different result: 54,154 tokens. This discrepancy is more significant than the actual number, as it exposes the flaw in the initial approach.

The original method involved running the same review task with multiple agents and attributing the difference to per-agent overhead. However, this method was flawed as it counted the same 54,154 context tokens multiple times, inflating the actual spawning cost. The cleaner measurement involved spawning a minimal subagent with a simple prompt, resulting in a single request containing 54,154 tokens in the agent's transcript file.

This revealed that the original 436,000 token cost was not the actual spawning cost but rather the cost multiplied by the number of times the context was thought to be paid for. The break-even analysis was redone, showing that delegating a read becomes cheaper than doing it in the main loop when the inline reading cost exceeds approximately 33,000 tokens.

This correction has important implications for measuring the cost of using Claude Code subagents, as it highlights the importance of accurate measurement methods to avoid overestimating the true cost.