In a "correction after rigorous testing," the author claimed that Codex's GPT-5.6 Luna Extra High delivers performance comparable to Terra High while being roughly 1.3x faster and 2.5x cheaper; however, without a published task suite or raw logs, this should be treated strictly as a routing hypothesis.
Entry point: Codex; the author discussed subagent quotas and Pro plan usage.
Comparison: GPT-5.6 Terra (High) vs. GPT-5.6 Luna (Extra High).
Test description: The author merely stated "correction after rigorous testing" without disclosing tasks, codebases, sample sizes, time windows, snapshots, tools, or billing breakdowns.
The original post did not provide full prompts or a downloadable test bundle; the reusable configuration hypothesis is to swap daily subagents from Terra High to Luna Extra High and compare them across the same categories of Codex tasks.
The author claims that performance is essentially identical.
The author claims that Luna is 1.3x faster than Terra.
The author claims that Luna is 2.5x cheaper than Terra; when used for Codex subagents, quota consumption is approximately 40% of Terra's.
For cost-sensitive Codex subagents where tasks are fallback-capable and validated by automated tests, Luna Extra High is worth evaluating as an A/B candidate against Terra High; Terra should still be reserved for failure retries or complex reasoning. This post does not prove equivalence across all tasks.
No published evaluation metrics for "identical performance," raw results, sample size, cost accounting methodology, or statistical uncertainty.
"2.5x less" likely refers to subscription quota/usage billing mechanics rather than raw API token pricing.
Codex product routing, quota limits, and model aliases are subject to change; results must be remeasured on current accounts and versions.
Select 20 independently verifiable Codex subtasks, holding the codebase, tools, context, and success criteria constant.
Randomly assign tasks to either Terra High or Luna Extra High, repeating for at least 3 rounds.
Record first-pass success rate, repair iterations, latency, tokens, account quota consumption, and manual quality scores.
Stratify and report results by task type; if Luna fails, log the fallback escalation rate to Terra and the total end-to-end cost.
GPT-5.6 Terra
