GPT-6.1 Terra cuts API prices 50% below GPT-6.1 Sol

GPT-6.1 Terra, the screen-reading and computer-use model OpenAI launched on October 26, costs $1 per million input tokens, $0.05 cached and $5 output in the API. That is exactly half of GPT-6.1 Sol on every rate, according to OpenAI’s announcement. If your team automates high-volume desktop work, such as keying invoices into accounting software, the chart data behind the post shows Terra tying GPT-6 Astra on Agents’ Last Exam at medium effort for about a sixteenth of the cost per task. For the hardest coding, OpenAI itself still points to Sol and Astra.

Agents’ Last Exam is OpenAI’s benchmark of professional computer-use tasks, from turning architectural drawings into a 3D model to logging an animated film shot by shot for a restoration team. At medium effort Terra and Astra both score 57.6%, at $0.29 against $4.82 per task. At max effort Terra leads 60.0% to 59.3%, for $1.01 against $7.24. The post itself quotes only the per-token gap, one-tenth of Astra’s price. OpenAI also reports a 15-point gain over GPT-5.6 Terra at medium effort for about a third of the cost, and says Terra completes tasks over three times faster than both Astra and GPT-6.1 Sol.

The invoice example in the post includes entering the details into accounting software, which is a computer-use step. For extraction alone, OpenAI’s guide to the GPT-6 family assigned jobs like pulling invoice fields to GPT-6 Luna in early October, and Luna lists at $0.10 input and $0.50 output.

Coding: where Sol stays ahead

On Terminal-Bench 4.0, OpenAI’s command-line coding test, Terra at max effort solves 50.3% of tasks, more than double GPT-5.6 Terra’s published 21.5%. But in the same chart, GPT-6.1 Sol at medium effort scores 52.7% for $1.02 per task, while Terra at max costs $1.84. For terminal coding, the model that is cheaper per token is the pricier route per task. Claude Sonnet 5.5 tops that chart at max effort with 70.6%.

Terminal-Bench-Science 0.1 has agents write code for research tasks such as tracking cell divisions in time-lapse video. Terra passes Sol there by just 1.3 points, and only at max effort: 58.3% at $4.15 per task against 57.0% at $5.47. At xhigh effort Sol leads, 53.7% to 43.4%. Claude Opus 5.5, labelled “w/ fallbacks” in the chart, scores 63.3% at $23.21, and GPT-6 Astra has the chart’s top score, 68.1% at $23.80.

Screens, images and errors

On OSWorld 2.1, which tests multistep tasks in desktop apps, Terra at medium effort scores 66.7% at $0.50 per task, slightly ahead of Sonnet 5.5 at medium (65.9%, $0.93). From high effort up, Sonnet leads, reaching 83.9% at max against Terra’s 74.2%. On MMMU-Pro, a visual reasoning test, Terra costs about half as much as Sol per item from low to xhigh effort. At max effort it costs 92% of Sol’s price and scores 0.8 points lower (85.5% against 86.3%).

OpenAI’s internal test of error-inducing prompts shows the factual error rate at low effort falling from 53.6% on GPT-5.6 Terra to 16.7%. Sol at low effort sits at 14.7%. At medium Terra is ahead, 10.1% to 12.0%, and at max the two cost the same $0.12 per prompt, with error rates of 4.3% and 4.4%.

Turning it on

In the API the model is gpt-6.1-terra. It accepts reasoning effort none, which GPT-6.1 Sol and Astra don’t offer, and has a June 1 knowledge cutoff, against April 30 for both of them. Cache writes cost $1.25 per million tokens for 30-minute retention and $2.50 for 12-hour retention. Prompts over 272K input tokens are billed at 2x input and 1.5x output for the whole request.

OpenAI says the same plan allowance in ChatGPT Work and Codex covers more tasks with Terra than with GPT-5.6 Terra, and substantially more than with Sol or Astra. The launch post lists Plus, Pro, Enterprise and Edu; the Codex changelog adds Business. The changelog also says Terra isn’t available in Chat and doesn’t support UAE inference residency, and that Enterprise access depends on workspace availability and admin controls. In the ChatGPT desktop app, you choose Terra from the model list in Advanced.