GPT-6.1 Sol and Terra Are Still Behind GPT-6 Astra

Beside its flagship, GPT-6 Astra, OpenAI sells cheaper models. Two of them are GPT-6.1 Sol, launched on Sep 29, and GPT-6.1 Terra, launched on Oct 26. Their launch posts chart them against Astra on eight benchmarks this site tracks. Sol scores below Astra on seven of its eight, and Terra on four of its five.

Each model matches Astra on one benchmark

Sol scores 75.2% on DeepSWE 1.1 at high effort, 1.1 points above Astra's 74.1%, in the Sol launch post. OpenAI wrote there that Sol "matches GPT-6 Astra at roughly one-fifth of the cost." Terra scores 60.0% on Agents' Last Exam v1 at maximum effort, 0.7 points above Astra's 59.3%, in the Terra launch post. OpenAI wrote that Terra "performs on par with GPT-6 Astra at only one-tenth its price per token."

Astra stays ahead on the rest

The largest gaps are on Terminal-Bench Science 0.1. Astra scores 68.1% there at maximum effort, Sol 57.0% and Terra 58.3%. OpenAI wrote in the Sol post that Astra "should be used for the most difficult scientific research tasks."

On Terminal-Bench 4.0, Terra's best is 50.3%, 7.9 points below Astra's 58.2%. The Terra post says: "For the most complex coding tasks, GPT-6.1 Sol and GPT-6 Astra remain the better choice." Sol is 5.3 points behind Astra on AutomationBench and 2.9 points behind on the offline set of OSWorld 2.1, where Terra is 6.7 points behind.

The Sol post also charts GDP.pdf, a benchmark this site does not track, where Sol scores just below Astra. Both posts chart OpenAI's internal test of factual errors on difficult prompts, where Astra makes the fewest errors of the three.

A fifth and a tenth of the price

Astra costs $10 per million input tokens and $50 per million output tokens on OpenAI's pricing page. Sol costs $2 and $10, a fifth of that. Terra costs $1 and $5, a tenth. Per task, the gap can be smaller. On Terminal-Bench Science at maximum effort, Sol costs $5.47 per task and Astra $23.80, OpenAI said in the Sol post.

How the scores were compared

Each row of the chart compares a model's best score at any reasoning effort with Astra's best on the same OpenAI chart. Astra's own launch post on Sep 3 gives it lower scores on three of these benchmarks: 57.9% on Terminal-Bench 4.0, 64.6% on Terminal-Bench Science 0.1 and 72.6% on the offline set of OSWorld 2.0, in its results tables. Sol and Terra are behind those figures too.

Terminal-Bench's own leaderboard lists Sol at 58.2% on Terminal-Bench 4.0, the same as Astra, above the 55.9% OpenAI charts for Sol in the Terra post. It does not list Terra. The benchmark makers' leaderboards for Terminal-Bench Science, DeepSWE and AutomationBench list Astra but neither 6.1 model, as of Nov 7.