Visual reasoning results on the MMMU-Pro benchmark show GPT-6.1 Terra outperforming its predecessor, GPT-5.6 Terra, while nearly matching top-tier frontier models like GPT-6 Astra and GPT-6.1 Sol across every reasoning configuration, at a fraction of the cost. At max effort, GPT-6.1 Terra achieved an 85.5% score at $0.012 per task, compared to GPT-5.6 Terra’s 81.7% at $0.079. Within the current family lineup, GPT-6 Luna scored 79.9% ($0.0021 per task), GPT-6.1 Sol hit 86.3% ($0.013 per task), and flagship GPT-6 Astra led at 86.7% ($0.065 per task). OpenAI did not include competitor model comparisons for this specific benchmark.
On the OSWorld 2.1 benchmark, which evaluates how AI agents navigate desktop applications to perform complex multi-step tasks, GPT-6.1 Terra scored 74.2% at max effort, costing $1.19 per task. That marks a substantial upgrade over GPT-5.6 Terra’s 57.1% score at $4.76 per task. However, it still trails Anthropic’s Claude Sonnet 5.5, which led the category at 83.9% with a cost of $5.74 per task.
OpenAI also claims to have invested heavily in reducing hallucinations and factual errors produced, and as per the internal benchmark of error-inducing prompts, the error rate at low reasoning effort fell from 53.6% with GPT-5.6 Terra to 16.7% with GPT-6.1 Terra, and at maximum effort, the error rate fell by 18 percentage points to 4.3%, nearly matching the rates for GPT-6 Astra and GPT-6.1 Sol. On cybersecurity, GPT-6.1 Terra meets the Critical threshold under OpenAI’s Preparedness Framework and retains the safeguards developed for GPT-6 Astra.
This story originally appeared on Computerworld
