GPT-6.1 Terra: $1/$5 pricing, computer use, benchmarks

OpenAI has introduced GPT-6.1 Terra, the mid-size model of its GPT-6.1 line, priced well below the larger ones. We read OpenAI’s announcement page on October 26, which says the model is available “starting today”. The system card it links to we did not read. Every benchmark figure below is OpenAI’s own, and the comparisons to Anthropic’s model are OpenAI’s choice of test and setting. We ran nothing.

What is it?

OpenAI calls Terra “a fast, efficient model that combines detailed visual reasoning and precise computer use alongside clearer, more natural writing”, a “significant step up” from GPT-5.6 Terra. Its examples are workday tasks: drafting proposals, interpreting technical drawings, filling out forms and editing videos. It is available to Plus, Pro, Enterprise and Edu users in ChatGPT Work and in Codex, and to developers as gpt-6.1-terra.

What does it cost?

Standard API pricing is $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens. OpenAI says these are 50% lower than GPT-5.6 Terra and GPT-6.1 Sol. In ChatGPT Work and Codex it says the same usage allowance covers more tasks than with GPT-5.6 Terra and “substantially more” than with GPT-6.1 Sol or GPT-6 Astra.

For scale, our earlier Opus 5.5 report lists $4 in and $20 out, so Terra’s list prices are one quarter of those (our arithmetic). That is a price comparison only. How this fits the wider run of cuts is in our price-cut piece.

What does OpenAI claim about performance?

• Computer use. On Agents’ Last Exam, a professional-task benchmark, OpenAI says Terra at medium reasoning effort scores 15 percentage points above GPT-5.6 Terra at about one third of the cost, and performs on par with GPT-6 Astra at one tenth of the price per token. At max effort it says Terra is “over three times as token efficient” as Claude Opus 5.5 and one quarter of its price per token, giving a cost per task about nine times lower. It also says Terra finishes tasks over three times faster than Astra and GPT-6.1 Sol.
• Visual reasoning and desktop tasks. Terra beats GPT-5.6 Terra at every reasoning setting on MMMU-Pro, and at medium effort on OSWorld 2.1 it passes GPT-5.6 Terra’s best score at a fraction of the cost. OpenAI gives these as charts without the numbers in the text.
• Factuality. On OpenAI’s internal benchmark of error-inducing prompts, the error rate at low effort fell from 53.6% to 16.7%, and at maximum effort fell 18 points to 4.3%.
• Coding. Terminal-Bench 4.0 at maximum effort: 50.3% of tasks, against GPT-5.6 Terra’s published 21.5%. OpenAI says GPT-6.1 Sol and GPT-6 Astra remain the better choice for the most complex coding. On Terminal-Bench-Science 0.1 at max reasoning it says Terra passes Sol at lower cost and comes within five points of Opus 5.5 at less than one fifth of the estimated cost.

What about safety?

OpenAI says GPT-6.1 Terra “meets the Critical threshold in cybersecurity under our Preparedness Framework” and keeps the protections developed for GPT-6 Astra. It also says Terra stays within scope and discloses limitations more honestly than GPT-5.6 Terra and resisted malicious instructions in its prompt-injection tests, and points to the system card for detail. We have not read the card, so we cannot say what the safeguards are or what the threshold means in practice.

What is still unknown?

• Independent results. All figures are OpenAI’s. We saw no third-party evaluation.
• The numbers behind the charts for Agents’ Last Exam, MMMU-Pro and OSWorld 2.1, and the settings used for the Opus 5.5 comparison.
• The system card contents, including the cyber evaluations behind the Critical threshold.
• Context window and rate limits. The announcement page we read does not state them.

What to keep in mind

These are our observations:

• A vendor comparison is a vendor choice. The Opus 5.5 figures come from tasks and settings OpenAI selected, and a cost-per-task claim depends on token use that you should measure on your own workload.
• Check cached versus uncached cost. A cached-input rate of $0.05 matters most for agents that re-read long contexts, as we discussed in the price-cut piece.
• Test the error and tool-use claims on your own tasks before moving a production workflow, since the error-rate benchmark is internal.

Editorial review: October 26, 2026 (AI). Prepared from OpenAI’s announcement page (read in full as page text; its charts show no numbers in the text). We did not read the system card, run the model or test any claim, and did not check OpenAI’s pricing page or the API documentation. The page may have changed since our reading.