The actual Codex Pareto frontier: Luna High → Terra Max → Sol Max — verified cost/performance across all 15 measured modes
SUMMARYA user revised a benchmark analysis of GPT-5.6 coding modes, comparing Luna High, Terra Max, and Sol Max across 15 measured configurations. The corrected results place Luna High at 1.00× quota with 56.8% CursorBench 3.2 performance, Terra Max at 3.52× quota with 64.9%, and Sol Max at 6.94× quota with 67.2%. The graph was rebuilt from a verified spreadsheet and official OpenAI documentation after earlier transcription and graphing errors were found.
I wanted to work out which combinations of GPT-5.6 model and reasoning effort provide the best coding performance for the quota they consume.
This graph plots all 15 quantitatively measured Luna, Terra and Sol configurations using:
- X-axis: average quota-equivalent cost per task, normalized to Luna High
- Y-axis: CursorBench 3.2 coding-agent performance
- Yellow circles: the recommended above-floor progression
The resulting three-step ladder is:
Key figures:
- Luna High: 1.00× quota, 56.8%
- Terra Max: 3.52× quota, 64.9%
- Sol Max: 6.94× quota, 67.2%
That means:
- one Terra Max task costs approximately 3.52 Luna High tasks
- one Sol Max task costs approximately 1.97 Terra Max tasks
- Sol Max is the frontier edge, but nearly doubles Terra Max’s cost for only a modest additional benchmark gain
Ultra is excluded from the quantitative graph because it is a separate multi-agent mode and there is currently no directly comparable same-harness score-and-cost result for it.
Correction and apology: I previously posted an earlier version of this graphic containing several data-transcription and graphing errors. That was my mistake, and the earlier version should be disregarded. Thanks to everyone who pointed out the problems—it helped me identify the errors and rebuild the analysis properly.
This replacement was constructed from a verified spreadsheet, cross-checked against the current CursorBench 3.2 results and official OpenAI model documentation, and generated programmatically from the underlying data rather than reconstructed by an image model.
Note: max thinking mode is a hidden option in Codex that you have to enable, which is strange, considering that it's the Pareto-leading mode for two of the steps.