GPT-6.1 Terra: OpenAI now sells cost per task, not price per token
OpenAI shipped GPT-6.1 Terra on Monday, a fast model built for visual reasoning and computer use, and the headline number on the launch post is not a benchmark score. It is cost per task. At max effort, OpenAI says Terra is "over three times as token efficient as Claude Opus 5.5, and one-fourth its price per token, resulting in a ~9x lower cost per task" on Agents' Last Exam.
The price list backs up the pitch: $1 per million input tokens, $0.05 per million cached input, $5 per million output, which OpenAI says is 50% below GPT-5.6 Terra and GPT-6.1 Sol. The context window is 1,050,000 tokens with 128,000 max output, and reasoning effort goes from none up to max. On Terminal-Bench 4.0 at max effort it solves 50.3% of tasks, more than double GPT-5.6 Terra's 21.5%. The low-effort hallucination rate fell from 53.6% to 16.7%. OpenAI also says Terra meets the Critical threshold for cybersecurity under its Preparedness Framework, which is a strange thing to put on the cheap model.
Read the fine print too. OpenAI itself says "for the most complex coding tasks, GPT-6.1 Sol and GPT-6 Astra remain the better choice." So Terra is not the smartest model. It is the model you put inside a loop that runs a thousand times. All the numbers are vendor numbers; an Artificial Analysis score of 48 is circulating on HN but we could not confirm it from the source.
The bigger shift is the framing. A year ago launches fought over leaderboard points. Now the three numbers buyers ask for are success rate, how often runs pass end to end, and dollars per finished task, and OpenAI chose the third as its lead. With cache reads already the bulk of coding-agent spend, a $0.05 cached-input price is aimed straight at agent workloads. Expect Anthropic and Google to answer with their own per-task math, not just a lower list price.
Launch post: https://openai.com/index/introducing-gpt-6-1-terra/ Model docs: https://developers.openai.com/api/docs/models/gpt-6.1-terra HN discussion: https://news.ycombinator.com/item?id=50248124
← Back to all articles
The price list backs up the pitch: $1 per million input tokens, $0.05 per million cached input, $5 per million output, which OpenAI says is 50% below GPT-5.6 Terra and GPT-6.1 Sol. The context window is 1,050,000 tokens with 128,000 max output, and reasoning effort goes from none up to max. On Terminal-Bench 4.0 at max effort it solves 50.3% of tasks, more than double GPT-5.6 Terra's 21.5%. The low-effort hallucination rate fell from 53.6% to 16.7%. OpenAI also says Terra meets the Critical threshold for cybersecurity under its Preparedness Framework, which is a strange thing to put on the cheap model.
Read the fine print too. OpenAI itself says "for the most complex coding tasks, GPT-6.1 Sol and GPT-6 Astra remain the better choice." So Terra is not the smartest model. It is the model you put inside a loop that runs a thousand times. All the numbers are vendor numbers; an Artificial Analysis score of 48 is circulating on HN but we could not confirm it from the source.
The bigger shift is the framing. A year ago launches fought over leaderboard points. Now the three numbers buyers ask for are success rate, how often runs pass end to end, and dollars per finished task, and OpenAI chose the third as its lead. With cache reads already the bulk of coding-agent spend, a $0.05 cached-input price is aimed straight at agent workloads. Expect Anthropic and Google to answer with their own per-task math, not just a lower list price.
Launch post: https://openai.com/index/introducing-gpt-6-1-terra/ Model docs: https://developers.openai.com/api/docs/models/gpt-6.1-terra HN discussion: https://news.ycombinator.com/item?id=50248124
Comments