OpenAI launches GPT-6.1 Terra, a cheaper mid-tier model built for computer use

OpenAI released GPT-6.1 Terra on October 26, a mid-tier reasoning model that succeeds GPT-5.6 Terra. The company positions it for everyday professional tasks that involve reading a screen and acting on it: filling in forms, editing videos, interpreting technical drawings, drafting proposals and automating repetitive desktop work.

Price and positioning

In the API, listed as gpt-6.1-terra, the model costs $1 per million input tokens, $5 per million output tokens and $0.05 per million cached input tokens, according to InfoWorld. That is half the price of the model it replaces. It sits between the low-cost GPT-6 Luna ($0.10 / $0.50) and the more capable GPT-6.1 Sol ($2 / $10) and GPT-6 Astra ($10 / $50). In ChatGPT, it is offered to Plus, Pro, Enterprise and Edu subscribers through ChatGPT Work and in Codex.

OpenAI still recommends Sol or Astra for the hardest coding work. BleepingComputer reports the company's claim that Terra finishes tasks more than three times faster than both of those models.

What the benchmarks show

All figures are published by OpenAI. On OSWorld 2.1, which measures desktop application tasks, Terra scores 74.2%, against 57.1% for GPT-5.6 Terra; Anthropic's Claude Sonnet 5.5 is listed higher, at 83.9%. On Agents' Last Exam, a test of professional tasks at maximum reasoning effort, Terra reaches 60.0%, slightly above GPT-6.1 Sol (58.6%) and GPT-6 Astra (59.3%) but below Claude Opus 5.5 (63.2%), at a much lower cost per task. On Terminal-Bench 4.0, it scores 50.3%, more than double its predecessor's 21.5%.

OpenAI also says that on an internal set of prompts designed to trigger mistakes, the error rate at low reasoning effort fell from 53.6% to 16.7%. Under its Preparedness Framework, the model is rated at the Critical level for cybersecurity and ships with the safeguards developed for GPT-6 Astra.

Open questions

Lower token prices do not automatically make agents cheaper to run. Analysts quoted by InfoWorld advise companies to test the model on their own workflows, tracking completion rates, accuracy and the human review time that agents still require. One argued that a few extra seconds of review per task can cancel out the inference savings. Independent evaluations of Terra had not been published at the time of writing, and OpenAI has not stated the model's context window in the material reviewed.