Article 328: GPT-6.1 Terra Is Half the Price and Beats Astra on Cost. It Still Loses to Claude Sonnet on Desktop Tasks

OpenAI shipped GPT-6.1 Terra on 26 October, replacing GPT-5.6 Terra in the middle of its own lineup - below GPT-6.1 Sol and GPT-6 Astra, above the cheap GPT-6 Luna. It is available today in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, on the API as gpt-6.1-terra, and same-day on Amazon Bedrock and Microsoft Foundry.

Standard API pricing: $1 per million input tokens, $0.05 cached, $5 per million output. That is half the price of GPT-5.6 Terra and GPT-6.1 Sol. The interesting part is not the discount. It is what OpenAI is claiming it gets for it.

The numbers, per task

Price per million tokens is the wrong unit for an agent. Cost per completed task is the right one, and OpenAI has started publishing its own:

  • Agents' Last Exam, max effort: Terra 60.0% at $1.01/task, versus GPT-6.1 Sol at 58.6% ($2.00) and GPT-6 Astra at 59.3% ($7.24). Claude Opus 5.5 scores higher at 63.2%, and costs $8.82 to get there.
  • MMMU-Pro, max effort: Terra 85.5% at $0.012/task, versus GPT-6 Astra at 86.7% ($0.065).
  • OSWorld 2.1, max effort: Terra 74.2% at $1.19/task, up from 57.1% ($4.76) on GPT-5.6 Terra.

Read the OSWorld line again. Terra jumped 17 points and cut cost by three quarters. And Claude Sonnet 5.5 still leads that benchmark at 83.9% - nine points clear, at $5.74 a task. So the honest summary is: Terra is the cheapest way to do competent desktop automation, not the best way to do hard desktop automation.

The number that actually matters

OpenAI's internal benchmark of error-inducing prompts: at low reasoning effort, the error rate fell from 53.6% to 16.7%. At low effort, the old model was wrong more often than right on prompts designed to trip it. That is the failure mode that makes people distrust agents, and it is the one they claim to have cut by more than two-thirds.

Terminal-Bench 4.0 also more than doubles, from 21.5% to 50.3% at max effort. Both are vendor benchmarks, published by the company selling the model, on prompts the company chose. Treat the direction as real and the exact figures as marketing.

Safety

Terra meets the Critical threshold in cybersecurity under OpenAI's Preparedness Framework. That is the same tier GPT-6 Astra crossed, and it is not a small thing: it means the cheap mid-tier model has the same cyber capability classification as the flagship did weeks ago. Cheap and capable spreads, whether or not anyone is ready.

Key Takeaway: The frontier stopped being where the money is. Terra scoring within one point of Astra on professional tasks at one-seventh the cost per task is the whole 2027 enterprise story in a pricing table - and the OSWorld number is the reminder that OpenAI is still not winning the capability it is now selling cheapest. Anthropic shipped a better computer-use model and priced it five times higher. That gap is the market. Somebody is going to fill it from the bottom.

Sources: OpenAI, "Introducing GPT-6.1 Terra," 26 October 2026; InfoWorld, "OpenAI bets on AI economics with lower-cost GPT-6.1 Terra," 27 October 2026; The Verge, 26 October 2026; AWS What's New, 26 October 2026