GPT-6.1 Terra Advances the Cost-Efficiency Frontier

GPT-6.1 Terra Advances the Cost-Efficiency Frontier

OpenAI has released GPT-6.1 Terra, the first Terra-class model since GPT-5.6. The model improves benchmark intelligence scores while reducing usage costs substantially.

Pricing is half or less than the rates for GPT-5.6 Terra. GPT-6.1 Terra costs $1 per million input tokens and $5 per million output tokens, compared with $2 and $12 for the previous model. Cache writes now cost $1.25 per million tokens, while cached input costs $0.05 per million tokens, down from $0.20.

Key findings

Stronger intelligence at lower cost

At maximum effort, GPT-6.1 Terra scores 6 points higher than GPT-5.6 Terra on the Artificial Analysis Intelligence Index, while the cost per task is less than one-third as high: $0.40 compared with $1.40.

The improvement is larger at lower effort levels, ranging from 7 to 13 points. Low and medium effort settings are Pareto-optimal for cost efficiency, while the other settings fall behind the cost-efficiency frontier.

At maximum effort, GPT-6.1 Terra scores 4 points below GPT-6.1 Sol and 10 points above GPT-6 Luna on the Intelligence Index.

Coding Agent Index performance

GPT-6.1 Terra scores 48 on the Artificial Analysis Coding Agent Index v2.0. This is 6 points below GPT-6.1 Sol, which scores 54, and 17 points above GPT-6 Luna, which scores 31.

The xhigh and low effort settings occupy the cost-efficiency frontier for this index. The remaining effort settings perform behind that frontier.

Lower hallucination rate

In AA-Omniscience, a private knowledge and hallucination benchmark, GPT-6.1 Terra gives an incorrect guess when it does not know the answer 38% of the time. GPT-5.6 Terra had a hallucination rate of 88%, representing a 50-point reduction.

The lower hallucination rate is accompanied by a 2-point increase in AA-Omniscience accuracy, indicating that the model answers more questions correctly while guessing less often.

Terminal-Bench improvement

GPT-6.1 Terra gains approximately 20 points on Terminal-Bench 4.0, a terminal-based agentic benchmark. At maximum effort, its performance rivals GPT-6.1 Sol.

Improvements across most other evaluations are smaller. For example, the model gains 5 points on GDP.pdf.

Evaluation summary

GPT-6.1 Terra scores 48 on the Artificial Analysis Coding Agent Index v2.0, compared with 54 for GPT-6.1 Sol and 31 for GPT-6 Luna. Its xhigh and low effort settings provide the strongest cost-efficiency results on this index.

The model also combines a 50-point reduction in AA-Omniscience hallucination rate with a 2-point accuracy increase. On Terminal-Bench 4.0, it improves by approximately 20 points and reaches performance comparable to GPT-6.1 Sol at maximum effort.

The Artificial Analysis Intelligence Index v4.3.2 provides a breakdown of the individual evaluation results.