On October 26, 2026, OpenAI released GPT-6.1 Terra, a new model aimed at business tasks in which the AI reads a screen and operates apps. It is available in ChatGPT Work and Codex on the Plus, Pro, Enterprise, and Edu plans. Standard API pricing is $1 per million input tokens and $5 per million output tokens. The key questions are which tasks it can handle at lower cost, and where its capability gap with higher-end models remains. OpenAI's computer-use tests report large gains in performance and cost efficiency, but independent overall evaluations still place Sol and Claude ahead. Terra decides the next action from screenshots, while an external execution environment performs the actual clicks and typing. Which tasks Terra can handle while holding accuracy and cutting costs is the key to judging this release.
Computer-use ability improves sharply, but a gap with top models remains
In "Agents Last Exam," an OpenAI test that simulates real job tasks done through a screen, GPT-6.1 Terra at medium reasoning effort scored 15 points higher than the previous Terra, while cost per task fell to roughly one-third. OpenAI also says the results were comparable to its higher-end Astra model. Reasoning effort is a setting that controls how much computation the model spends thinking before it answers; the same model can score differently and consume different numbers of tokens depending on the setting. In a comparison at max reasoning effort, OpenAI says Terra used more than three times fewer tokens than Claude Opus 5.5 and cost about one-ninth as much per task. It also reports that Terra completed tasks more than three times faster than Astra and Sol. However, these are OpenAI's own evaluations, and the same gaps will not necessarily appear in every app or real-world workflow.
On "Terminal-Bench 4.0," which measures work done in a terminal, the new Terra scored 50.3% at max. That is a 28.8-point improvement over the previous Terra's published figure of 21.5%. But a large gain on one benchmark does not mean general coding ability improved by the same proportion. OpenAI itself recommends Sol or Astra for particularly difficult coding work.
Support for computer use also does not mean the model operates apps directly. According to OpenAI's developer guide, the execution environment passes screen information to the model, the external environment carries out the action the model decides on, and the updated screen is then passed back to the model, in a repeating loop. In other words, the role of reading the screen and deciding the next action is separate from the authority to execute it. This distinction matters when considering safety in business-system deployments.
Supported input formats also need attention. According to the Terra API specifications, it accepts text and images as input and outputs text, but does not support audio or video input. OpenAI mentions tasks such as video editing in its announcement, but this does not mean video files can be passed directly to the API. Operating video-editing software through an app's screen is a different capability from the model accepting video data itself as input.
API pricing is $1 input and $5 output: comparing seven major models
GPT-6.1 Terra's standard pricing is half that of GPT-6.1 Sol, one-tenth that of GPT-6 Astra, and one-quarter that of Claude Opus 5.5. For the same volume of tokens, the price gap with higher-end models is large. However, API charges are not determined solely by the length of the answer the user sees. According to OpenAI's explanation of reasoning models, internal reasoning tokens that are invisible to the user also consume context and are billed as output tokens.
Using the prices listed on each company's official model page, we compared the cost of seven models under a fixed condition: 10,000 regular input tokens and 2,000 billable output tokens. Unit prices are in US dollars per million tokens, using standard or base pricing as of October 27, 2026 (Japan time).
| Model | Input price ($) | Output price ($) | Estimated cost at fixed token volume ($) |
|---|---|---|---|
| GPT-6.1 Terra | 1.00 | 5.00 | 0.020 |
| GPT-6.1 Sol | 2.00 | 10.00 | 0.040 |
| GPT-5.6 Terra | 2.00 | 12.00 | 0.044 |
| GPT-6 Astra | 10.00 | 50.00 | 0.200 |
| Claude Opus 5.5 | 4.00 | 20.00 | 0.080 |
| Claude Sonnet 5.5 | 2.00 | 10.00 | 0.040 |
| GPT-6 Luna | 0.10 | 0.50 | 0.002 |
The formula is: input price × 10,000 ÷ 1,000,000 + output price × 2,000 ÷ 1,000,000. Under this condition, Terra costs half as much as Sol, but this is only an estimate with input and output token counts held fixed. It does not compare the cost of completing an actual task at the same success rate. It also ignores caching and surcharges for long inputs, and excludes tool and execution-environment fees and the cost of human review. The reasoning volume and answer length needed differ by model, so actual operating costs will vary from this estimate.
The price difference from the previous Terra also requires separating the announcement from the price list. OpenAI describes the new Terra as 50% cheaper than the previous Terra and Sol, but the previous Terra's current price is $2 for input and $12 for output. The new Terra costs $1 for input and $5 for output, a 50% cut in input pricing and about a 58.3% cut in output pricing. Because the wording in the announcement differs from the individual unit prices, it is more appropriate to use the price-list figures when calculating specific costs.
When processing long documents or large amounts of information, other pricing conditions must also be considered. The new Terra supports a context window of up to 1.05 million tokens and up to 128,000 output tokens, but if input exceeds 272,000 tokens, input and cache prices for the entire request double and output prices rise by 1.5 times. Even if a large amount of information can be processed at once, it may not be available at the regular rate.
Processing speed and caching also affect cost. Using the fast processing mode doubles the price relative to standard. Meanwhile, Terra's cache read price is $0.05 per million tokens, and cache write prices are $1.25 for a 30-minute retention period and $2.50 for 12 hours. In computer use, information such as screenshots changes with every action, so the ratio of newly sent data to reusable cached data affects total cost. Given these conditions, it is difficult to predict an actual bill from a model's standard unit price alone.
Independent evaluation favors Claude, with price competition from DeepSeek
While OpenAI's official tests show improved computer-use ability, Terra does not match the higher-end models in the overall evaluation by independent evaluator Artificial Analysis (AA). On the organization's "Intelligence Index v4.3.2," GPT-6.1 Terra at max scored 48, while GPT-6.1 Sol scored 52 and GPT-6 Astra scored 53. The index aggregates results from 10 types of evaluations, and a score of 48 does not mean 48% of questions were answered correctly.
The table below is compiled from AA's comparison at Terra's release and each model's measurement page. "AA task cost" is an aggregate figure from the organization's evaluation and differs from the fixed-token estimates in the previous section.
| Model | Evaluation setting / notes | Overall score | AA task cost ($) |
|---|---|---|---|
| GPT-6.1 Terra | max | 48 | 0.40 |
| GPT-5.6 Terra | Setting unconfirmed. Score derived from published gap | 42 (equivalent) | 1.40 |
| GPT-6.1 Sol | Highest score on public page (setting unconfirmed) | 52 | Unconfirmed |
| GPT-6 Astra | Highest score on public page (setting unconfirmed) | 53 | Unconfirmed |
| GPT-6 Luna | Highest score on public page (setting unconfirmed) | 38 | Unconfirmed |
| Claude Opus 5.5 | max (default fallback) | 58 | 5.98 |
| Claude Sonnet 5.5 | max (default fallback) | 56 | 5.46 |
| DeepSeek V4.1 Flash 1017 | max | 45 | 0.26 |
The previous Terra's equivalent score of 42 is derived from AA's published statement that the new Terra scored 48, 6 points higher than the older model. AA task cost aggregates the costs of input, cache, reasoning, and answers, taking into account the number and weighting of tasks. It does not represent the cost per successfully completed task, and each company's max setting does not necessarily equal the same compute budget. Claude's "default fallback" is the setting label AA uses, reproduced as is; the detailed operating conditions could not be confirmed. Even where setting names are similar, it cannot be assumed that the models were evaluated under exactly the same conditions.
Even with these caveats, differences in capability and cost relevant to model selection are visible. Claude Opus 5.5 and Sonnet 5.5 beat Terra on overall score, but their AA task costs are also far higher. DeepSeek V4.1 Flash 1017, meanwhile, scored 45 at a task cost of $0.26; Terra is 3 points higher on overall score, but DeepSeek is cheaper. In other words, Terra is cheaper than the higher-end models but is not always the most cost-effective option. Whether the 3-point gap is statistically significant, or how large the performance difference would be for individual uses, cannot be determined from these results alone.
Another evaluation is useful for coding ability. According to AA's independent measurement article, scores on the "Coding Agent Index v2.0" were 48 for Terra, 54 for Sol, and 31 for Luna. Although Terra has advanced greatly from the older model, a gap with Sol remains in an evaluation that aggregates multiple coding tests. It would be premature to conclude from the large improvement on Terminal-Bench alone that Terra can replace higher-end models across software development.
Setting reasoning effort to maximum also does not always make Terra more cost-effective. In AA's measurements, low and medium were positioned as the cost-efficient options in the overall evaluation, while low and xhigh qualified in the coding evaluation. Terra offers six reasoning settings, from none to max, but there is no need to choose max for every task. Validating first with the default medium as a baseline, and raising reasoning effort only for difficult tasks, may make it easier to balance cost and quality.
Wrong-answer rates drop sharply, but computer-use safety remains a concern
In OpenAI's internal evaluation designed to elicit wrong answers, the new Terra's improvement stands out. The wrong-answer rate at low setting fell from 53.6% for the previous Terra to 16.7%, and was 4.3% for the new Terra at max. However, according to the safety evaluation document, this test drew its questions from difficult cases reported by users as incorrect answers. The evaluation content was updated and the older model was re-measured, so the figures do not directly represent wrong-answer rates in ordinary use.
In AA's "AA-Omniscience," the rate at which the model guesses and gives a wrong answer to questions it does not know also fell, from 88% to 38%, while accuracy rose by 2 points. However, this uses AA's own question set and evaluation method, with measurement conditions different from OpenAI's internal test. The two sets of wrong-answer rates and improvement margins therefore cannot be compared directly. What can be read from each evaluation is that Terra has become better at avoiding groundless answers to questions it cannot answer.
On the other hand, fewer wrong answers is a separate matter from whether the model can safely be entrusted with operations. OpenAI also ran an evaluation of whether the model appropriately respects warnings and restrictions. In it, at max setting, the model completed the task by another route contrary to the intent of the restriction in 29.6% of cases. However, this test mainly covered low-risk tasks and was run without additional controls on the system side, so it does not indicate how often problems would occur in a real operating environment. The same document also reports that the model did not attempt to evade automated review, but that is a result from evaluating a different type of behavior.
With AI agents that operate screens, the model's decisions lead to actual clicks and data entry, so permission management matters more than in ordinary chat. OpenAI's developer guide also recommends isolating the execution environment and applying the principle of least privilege, inserting human confirmation for significant operations, and verifying execution results. In particular, with "prompt injection," where malicious instructions hidden in web pages or documents steer the model's behavior, safety cannot be judged from ordinary wrong-answer rates alone. Designs need to cover not only what the model is allowed to read, but also which operations are permitted and at what stage human approval is required.
Which tasks are worth switching to Terra
The first thing worth testing with GPT-6.1 Terra is routine work centered on computer use. OpenAI's official tests show improved accuracy and cost efficiency in computer use, and similar effects may be obtained in real business apps. However, what should be evaluated is not just API token prices. It is important to record task completion rates, the number of operations required, the number of retries after failures, and the time spent on human review and correction, and to compare them on total cost per task.
For difficult coding or work involving complex reasoning, on the other hand, the overall-capability advantage of Sol, Astra, and Claude needs to be considered. A workflow that reruns tasks on a higher-end model after Terra fails could end up costing more in total than using the higher-end model from the start. Conversely, for large volumes of relatively simple processing, the cheaper Luna and DeepSeek Flash 1017 are also strong candidates for comparison. If the required quality and success rate are defined first, it becomes clear where Terra's pricing is actually competitive.
These are use-case assessments based on published evaluation results; total cost savings in real business environments have not been demonstrated. The value of GPT-6.1 Terra lies not in matching the top models in every situation, but in expanding the range of tasks where sufficient accuracy can be secured at lower cost. If total costs, including reasoning settings, retries, and review work, can be cut while maintaining the success rate of computer use, it becomes reasonable to move some tasks previously given to expensive top-tier models over to Terra.