Contents
OpenAI released GPT-6.1 Terra on October 26, 2026, with improved visual reasoning and computer use at half the standard API pricing of GPT-5.6 Terra. It is worth trying for repetitive tasks across multiple applications, though OpenAI continues to recommend GPT-6.1 Sol and GPT-6 Astra for more complex coding.
Key points#
- GPT-6.1 Terra is available through the API and to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex.
- Standard API pricing per million tokens is $1 for input, $0.05 for cached input, and $5 for output.
- According to OpenAI, at the medium reasoning setting, Terra outperformed GPT-5.6 Terra by 15 percentage points on Agents’ Last Exam at roughly one-third the cost.
- Inputs exceeding 272,000 tokens trigger higher pricing for the entire request, making context size a key cost factor.
What Terra improves#
GPT-6.1 Terra focuses on tasks that involve understanding screens and acting on their contents. OpenAI cites invoice processing as an example: reading documents and entering details into accounting software.
This makes it suited to workflows requiring both visual interpretation and precise actions. Reading an invoice correctly is only part of the task; selecting the right fields, entering values, and navigating the application are also necessary.
OpenAI reports clearer, more natural writing and improved coding performance. This could broaden its uses in everyday work, but does not establish it as the best choice for every task.
What the benchmarks show#
On Agents’ Last Exam, which evaluates agents using business software, OpenAI reports that Terra at the medium reasoning setting outperformed GPT-5.6 Terra by 15 percentage points at roughly one-third the cost. The evaluation includes tasks such as creating 3D models from architectural drawings.
At the maximum reasoning setting, OpenAI says Terra matched GPT-6 Astra on the same benchmark at one-tenth the price per token. It also completed tasks more than three times as fast on average as GPT-6 Astra and GPT-6.1 Sol. These are benchmark results reported by OpenAI; they do not guarantee the same cost savings or speed gains in individual workflows.
On MMMU-Pro, a visual reasoning benchmark, Terra also outperformed GPT-5.6 Terra at every reasoning setting tested. At the maximum reasoning setting, results also improved on Terminal-Bench 4.0, though OpenAI continues to recommend Sol and Astra for more complex coding.
Access and settings#
API users can specify the model ID gpt-6.1-terra. According to OpenAI, it is available in ChatGPT Work and Codex on Plus, Pro, Enterprise, and Edu plans, allowing users to complete more work within existing usage limits than with the comparison models.
The context window is 1.05 million tokens, with a maximum input of 922,000 tokens and a maximum output of 128,000 tokens. Supported reasoning settings are none, low, medium, high, xhigh, and max, with medium as the default.
Start evaluations at the medium reasoning setting. Increase it if tasks fail because they require deeper reasoning, then compare the quality of completed work, elapsed time, and total cost. The maximum reasoning setting should deliver results that justify the additional processing.
When API costs rise#
Standard pricing per million tokens is as follows.
- Input: $1.
- Cached input: $0.05.
- Output: $5.
- Fast mode: twice the standard rates.
Pay attention to the pricing threshold for long contexts. When input exceeds 272,000 tokens, input and cached input rates double, and output rates increase by 50%. These higher rates apply to the entire request, not just the tokens above the threshold.
For example, a standard request with 300,000 uncached input tokens and 10,000 output tokens costs $0.60 for input and $0.075 for output, totaling $0.675 before tool fees. Some tools incur separate charges, so token pricing alone does not capture the cost of the entire workflow.
Reliability still needs checking#
According to OpenAI, in an evaluation based on conversations where users reported factual errors, Terra at the low reasoning setting had roughly one-third the factual error rate of GPT-5.6 Terra. This is a result from a specific evaluation, not a general error rate across all responses.
The company also reports improvements in staying within task scope, candidly disclosing limitations, and resisting the prompt injection attacks tested. Detailed methods and results are provided in the system card.
For computer use, evaluate completed work rather than persuasive explanations. Check whether it entered the correct amount, used the intended record, and stopped when information was missing. Continue requiring human review for actions that finalize payments or change important records until your own testing supports broader use.
Next steps#
Run one representative invoice-processing task with GPT-6.1 Terra at the medium reasoning setting, and compare field-entry accuracy, completion time, and total API cost with your current workflow.
Source: https://gihyo.jp/article/2026/10/gpt-6-1-terra
Contents
