GPT-6.1 Terra Is Here: Prices Cut in Half, Hallucination Rate Plummets, but There Are Too Many Choices
OpenAI has released GPT-6.1 Terra, cutting output pricing to $5 per million tokens while intelligence actually increased. At low/medium effort, the price-performance ratio is maxed out, and the hallucination rate has dropped from 88% to 38%. But with more and more combinations of models and effort levels, choice paralysis is starting to spread.
OpenAI today released GPT-6.1 Terra. It is the first Terra-class model after GPT-5.6, with a very straightforward positioning: lower costs and raise intelligence.
Price is the biggest change. Input is $1 per million tokens and output is $5, compared with $2/$12 for GPT-5.6 Terra, so output price is halved. Cache writes drop to $1.25, and cache reads fall from $0.20 to $0.05.
On performance, Artificial Analysis's Intelligence Index test shows that at max effort, GPT-6.1 Terra scores 6 points higher than GPT-5.6 Terra, while cost per task drops from $1.40 to $0.40—less than one-third of the original. At low and medium effort, the gains are even larger, at 7 to 13 points. Cost efficiency at these two settings is already on the Pareto frontier. The max effort score is 4 points lower than GPT-6.1 Sol and 10 points higher than GPT-6 Luna.
In coding, Terra scores 48 on the Coding Agent Index v2.0, 6 points lower than Sol and 17 points higher than Luna. The xhigh and low effort settings are on the cost-efficiency frontier.
The most notable improvement is the hallucination rate. On the AA-Omniscience benchmark, when Terra faces questions it does not know, the proportion of guessed answers drops from 88% to 38%. Accuracy also rises by 2 percentage points. This means the model simultaneously makes things up less and gets more answers right.
On the terminal-agent benchmark Terminal-Bench 4.0, Terra jumps about 20 points, approaching Sol's max effort level. Gains on other evaluations are smaller; GDP.pdf rises by 5 points.
The community reaction has split into two camps. One camp focuses on price, with a netizen saying, "Output at $5, and my API bill has completely changed." The other camp is disappointed; one user said, "The intelligence gain isn't enough; I'll stick with Astra." Others complain about too many choices: OpenAI and Anthropic together have roughly 40 model/effort combinations, which is already overwhelming.
Terra did not surpass Sol, but it has taken “good enough and cheap” to a new level. When cost is no longer a barrier, the way AI gets deployed may change.
发布时间: 2026-10-27 03:48