OpenAI’s new middle model scores 3.7 points below GPT-6.1 Sol on Artificial Analysis’s Intelligence Index, and costs 56% as much to run each test task.
By ModelCompare ·
OpenAI released GPT-6.1 Terra on Monday, in ChatGPT, Codex and the API. It sits in the middle of OpenAI’s four tiers: above GPT-6 Luna ($0.10 per million tokens in, $0.50 out), below GPT-6.1 Sol ($2 and $10) and GPT-6 Astra ($10 and $50). Terra lists at $1 in and $5 out: exactly half of Sol.
Artificial Analysis measured that the day Terra came out. Running its Intelligence Index, a set of tests in reasoning, knowledge, coding and agent work, Terra spent $0.40 per task and scored 48.1. Sol spent $0.72 and scored 51.8. So the discount holds: Terra does each task for 56% of Sol’s cost, 3.7 points behind.
Score against cost per task, every model we track that has both, on 26 OctoberUp is smarter, left is cheaper. The dashed line joins the models nothing beats on both.
Source: Artificial Analysis Intelligence Index and its cost to run, per task, in US dollars on a log scale. Each model at the setting AA lists for it, usually its highest.
That puts Terra on the line of best value. No model we track scores higher for less. The cheapest model that scores at least as well is GPT-6.1 Sol, at $0.72 a task. Below Terra’s price, the best score belongs to MiMo V2.6 Pro: 46.3, at $0.13.
Within OpenAI’s own line, Terra fills a wide gap. GPT-6 Luna costs $0.07 a task but scores 38.1. GPT-6 Astra, the top tier, scores 52.7 at $3.26 a task, 8 times Terra’s cost.
What OpenAI claims
OpenAI pitches Terra for computer use, tool calls and routine coding. It says Terra finishes agent tasks more than 3 times faster than Sol, and that at its lowest reasoning setting it made errors on 16.7% of a set of tricky prompts, down from 53.6% for the Terra before it. Nobody has checked those numbers independently yet.
Try your own prompt on both and see what each answer costs.
Scores and costs per task are Artificial Analysis’s, as of 26 October, for every model we track that had both. AA runs each model at the setting it names in brackets on its site (GPT-6.1 Terra (Max), GPT-6.1 Sol (Max)); cost per task is its total bill for the Intelligence Index divided by the number of tasks, at list prices. Prices are OpenAI’s list prices per million tokens, for prompts up to 272,000 tokens.