Benchmarking GPT-5.5 against three GPT-5.6 models: Sol, Terra, and Luna, inside real VDR extraction workflows

On accuracy, Sol is the clear leader: GPT 5.6 Sol scored 93.60%, followed by Luna at 86.40%, Terra at 86.00%, and GPT 5.5 at 85.60%. Sol’s eight-percentage-point advantage over GPT-5.5 is substantial, but the broader results show that each model balances accuracy, depth, and speed differently.

Sol was the most exhaustive and presentation-heavy. It averaged 363.7 words per response, 2% more than GPT-5.5’s 355.1. In our separate qualitative review, Sol used a table in all 20 responses and did not withhold an answer because of limited source coverage. GPT-5.5 remained balanced and consistent, with more audit-friendly calculation notes, while the GPT-5.6 configurations were generally more answer-first and included less process narration.

The metadata makes those trade-offs especially clear. Compared with GPT-5.5:

• Sol used 17% more output tokens and made 62% more tool calls while responding 11% faster.
• Luna used 13% fewer tokens and 26% fewer tool calls while responding 43% faster.
• Terra reduced both response time and word count by about 20%, despite making 53% more tool calls.

Our takeaway is that Sol maximized accuracy and depth, Luna had the leanest resource profile, and Terra favored concise output.

Contributing authors: Maya Boeye, Jeondo Lee, Steiner Williams, Jennifer Venus