GPT-6.1 Terra Makes OpenAI’s Midrange a Value Champion
Set Trending Topics as a preferred source on Google.
OpenAI has just released GPT-6.1 Terra, the first Terra-class model since GPT-5.6. According to Artificial Analysis, it pushes the cost efficiency frontier further: more intelligence for less than a third of the cost per task.
Terra is the middle performance tier in OpenAI’s model family. With GPT-5.6, the company introduced three tiers last summer: Sol, Terra and Luna. Terra sits between the stronger Sol and the weaker Luna and stands for the balanced middle of performance and price. The GPT-6 generation had no Terra so far; it launched with Astra as the flagship, plus Sol and Luna.
Half the Price, More Points
Token prices are $1 per million input tokens and $5 per million output tokens (about €0.89 and €4.45). The predecessor, GPT-5.6 Terra, cost $2 and $12 (about €1.78 and €10.68). Caching also gets cheaper: cache writes drop to $1.25 (about €1.11), and cached input falls to $0.05 (about €0.04) from $0.20 (about €0.18).
In the Artificial Analysis Intelligence Index, GPT-6.1 Terra gains 6 points over GPT-5.6 Terra at maximum reasoning effort. The cost per task falls from $1.40 to $0.40 (about €1.25 to €0.36). At lower effort levels, the gain is even larger, between 7 and 13 points.
Within OpenAI’s own lineup, Terra at its highest setting sits 4 points behind GPT-6.1 Sol and 10 points ahead of GPT-6 Luna. Sol launched at the end of September at $2 and $10 per million tokens (about €1.78 and €8.90), so Terra now costs half as much as Sol for both input and output.
Anyone watching costs should keep an eye on the effort levels: Artificial Analysis rates low and medium as Pareto-optimal, meaning they offer the best balance of intelligence and cost. The other levels sit behind that frontier.
Fewer Hallucinations, Better Coding Agents
The most striking jump concerns reliability. In AA-Omniscience, a private knowledge and hallucination test, GPT-6.1 Terra guesses in 38 percent of cases when it does not know the answer. Its predecessor did so 88 percent of the time. Accuracy rises by 2 points at the same time, so the model makes things up less often and still answers more questions correctly.
On the newly released Artificial Analysis Coding Agent Index v2.0, Terra scores 48 points. Sol reaches 54 and Luna 31. The xhigh and low settings sit on the cost efficiency frontier here. On Terminal-Bench 4.0, a terminal-based agent benchmark, Terra gains about 20 points and now rivals GPT-6.1 Sol at maximum effort. Gains on most other tests are smaller, such as 5 points on GDP.pdf.
For developers and companies running agents at scale, these are the levers that matter: lower cost per task and fewer fabrications.
The Competition Is Not Standing Still
OpenAI is under pressure. Anthropic recently overtook its rival in the rankings with the smaller Sonnet 5.5. Open models are also climbing: Moonshot AI’s Kimi K3 had already made waves, and its successor Kimi K3.1 Pro launches in fourth place on the Intelligence Index, according to Artificial Analysis, and is set to lead the open weights models.
Where the Limits Lie
The figures come from Artificial Analysis, an independent benchmarking firm, and they also show weaknesses. The hallucination rate fell from 88 to 38 percent, yet the model still guesses in more than a third of cases when it lacks an answer. On the Coding Agent Index, Terra trails Sol by 6 points, and by 4 points on the overall index. Only some of the effort levels sit on the cost efficiency frontier: at maximum effort, Terra is not Pareto-optimal, according to Artificial Analysis.
