Products & Services · 2026-10-28 · Team LATENT

What does one computer-use task cost on GPT-6.1 Terra, and what do you pay for the gap to Astra?

GPT-6.1 Terra's API rates are one tenth of GPT-6 Astra's. Using every data point in OpenAI's launch charts, we read the cost of one computer-use task at each reasoning effort, then estimate what you pay for Astra's higher scores and what screenshots cost.

Read an incoming invoice and enter its line items into accounting software. OpenAI cited this task in the GPT-6.1 Terra launch as an example of high-volume work to automate. If you hand it to an AI that looks at the screen and drives the mouse and keyboard, how much does each invoice cost?

On October 26, 2026, OpenAI released GPT-6.1 Terra, a fast model for visual reasoning and computer use. It is available in ChatGPT Work, Codex, and the API. Standard API pricing is $1 per million input tokens and $5 per million output tokens, one tenth of the flagship GPT-6 Astra.

This article reads the cost of one task at each reasoning effort, using every data point in the charts on OpenAI's launch page. It then estimates what you pay for Astra's higher scores and what screenshots cost when you run the model yourself. Specifications and pricing were checked on October 27. Benchmark costs and scores are OpenAI's published figures, and the estimates are LATENT's calculations from those figures and the price list.

OpenAI announcement / System card addendum / API pricing

Key takeaways

Raising Terra's reasoning effort from low to max multiplies the cost of one task by 3.3 on OSWorld 2.1 and by 6.3 on Agents' Last Exam. The per-token rate alone does not set the cost of a task.
On OSWorld 2.1, GPT-6.1 Sol, at twice the per-token rate, beats Terra at about the same cost. Sol at medium scores 73.5% for $0.65, while Terra at xhigh scores 71.3% for $0.62.
On Agents' Last Exam, Terra and Astra at medium tie at 57.6%, and Terra costs about one seventeenth as much. Astra leads on OSWorld 2.1, Terminal-Bench 4.0, and robustness to indirect prompt injection.

Rates are one tenth of Astra's, but reasoning effort moves the cost of a task by 3 to 6 times

Per million tokens, Terra costs $1 for input, $0.05 for cached input, and $5 for output. The previous GPT-5.6 Terra cost $2 for input and $12 for output. GPT-6.1 Sol costs $2 and $10, exactly twice Terra. GPT-6 Astra costs $10 and $50, ten times Terra. Anthropic's Claude Opus 5.5, which appears in the charts, costs $4 and $20, four times Terra.

These rates alone do not determine what a task costs. Terra's API offers six reasoning effort levels, from none to max. The launch charts show the cost per task and the score for the five levels other than none: low, medium, high, xhigh, and max.

Two of the evaluations cover computer use. OSWorld 2.1 measures multi-step tasks carried out in desktop applications, and OpenAI gives building a 3D model from a technical drawing as an example. Agents' Last Exam, co-led by UC Berkeley RDI and others, measures long professional workflows. Its tasks include building a 3D model from architectural drawings and reviewing an animated film shot by shot.

On OSWorld 2.1, Terra scored 53.8% for $0.36 at low and 74.2% for $1.19 at max. The cost rises 3.3 times and the score rises 20.4 points. On Agents' Last Exam, Terra scored 55.9% for $0.16 at low and 60.0% for $1.01 at max. The cost rises 6.3 times, but the score rises only 4.1 points. What higher reasoning effort buys differs widely by task type.

The chart values also confirm the comparisons in OpenAI's text. On Agents' Last Exam at medium, Terra scored 57.6% for $0.29, and GPT-5.6 Terra scored 42.6% for $0.84. That is 15.0 points higher for about one third of the cost. At max, Claude Opus 5.5 scored 63.2% for $8.82, and Terra scored 60.0% for $1.01. The cost gap is about 8.7 times, matching the "about one ninth" in OpenAI's text.

OpenAI announcement / GPT-6.1 Terra model page / GPT-5.6 Terra model page / Claude pricing

On OSWorld, Sol beats Terra at the same cost despite twice the rate

The table lays out the OSWorld 2.1 chart by reasoning effort. Each cell shows the cost per task and the score.

Reasoning effort	GPT-6.1 Terra	GPT-6.1 Sol	GPT-6 Astra
low	$0.36 / 53.8%	$0.38 / 60.9%	$2.32 / 63.3%
medium	$0.50 / 66.7%	$0.65 / 73.5%	$4.37 / 75.4%
high	$0.51 / 68.7%	$0.90 / 75.5%	$5.76 / 78.3%
xhigh	$0.62 / 71.3%	$0.95 / 75.9%	$6.60 / 79.0%
max	$1.19 / 74.2%	$1.15 / 78.0%	$8.17 / 80.9%

At low, Sol costs only $0.02 more than Terra and scores 7.1 points higher. Sol at medium scores 73.5% for $0.65, above Terra at xhigh with 71.3% for $0.62. At max, Sol is both cheaper and higher scoring. Whenever a Sol level and a Terra level cost about the same, Sol scores higher.

Sol's API rates are twice Terra's for input, cached input, and output alike. If a model with twice the rate costs about the same per task, it must use about half as many tokens per OSWorld 2.1 task as Terra. Choosing the model with half the rate does not guarantee half the cost per task.

On Agents' Last Exam, Terra has the edge instead. Terra at low scores 55.9% for $0.16, beating Sol at low (52.4% for $0.70) on both cost and score. Terra at xhigh scored 58.5% for $0.66, and Sol at max scored 58.6% for $2.00. Even among computer-use evaluations, which model is cheaper flips with the task.

The highest score on the OSWorld 2.1 chart is Claude Sonnet 5.5 at max, 83.9% for $5.74. Sonnet 5.5 at xhigh scored 81.0% for $2.22. That is about the same as Astra at max ($8.17, 80.9%) for about a quarter of the cost. All of these are OpenAI's measurements.

The scores need careful reading. OSWorld 2.1's grading code returns a score from 0 to 1 for each task, a weighted sum over several criteria. OpenAI's chart does not say whether its score is an average including this partial credit or the share of tasks that met every criterion. For that reason, this article does not call the scores "success rates." The public repository includes a list of 82 tasks that run without internet access. The announcement does not say whether OpenAI's "offline set" is that same list.

OSWorld 2.1 release notes / OSWorld-V2 result format

What you pay per point of Astra's lead

The table compares Astra and Terra on the evaluations in the announcement. The cost ratio is Astra's cost per task divided by Terra's.

Evaluation	Reasoning effort	GPT-6.1 Terra	GPT-6 Astra	Cost ratio
Agents' Last Exam v1	medium	$0.29 / 57.6%	$4.82 / 57.6%	16.6x
Agents' Last Exam v1	max	$1.01 / 60.0%	$7.24 / 59.3%	7.2x
OSWorld 2.1	medium	$0.50 / 66.7%	$4.37 / 75.4%	8.7x
OSWorld 2.1	max	$1.19 / 74.2%	$8.17 / 80.9%	6.9x
Terminal-Bench 4.0	max	$1.84 / 50.3%	$10.34 / 58.2%	5.6x
MMMU-Pro	medium	$0.0041 / 84.6%	$0.039 / 85.3%	9.5x
Share of answers with factual errors (lower is better)	medium	$0.029 / 10.1%	$0.29 / 9.7%	10.0x
Robustness to indirect prompt injection	not stated	98.770%	99.789%	not stated

On Agents' Last Exam, the extra you pay for Astra does not show up in the score. The two tie at 57.6% at medium, and at max Terra is 0.7 points higher. For work resembling this evaluation, Terra is enough.

The gap appears on OSWorld 2.1. At medium, Astra scores 8.7 points higher than Terra and costs $3.87 more per task. That works out to about $0.44 per point for each task. Across 100 tasks, the extra comes to $387, and the summed score difference equals 8.7 tasks' worth of work. That is about $44.5 per task's worth. If having a person finish the remaining work costs more than $44.5 per task, that is a reason to choose Astra at medium.

The amount changes sharply, though, with what you compare against. Terra at max scores 74.2% for $1.19, within 1.2 points of Astra at medium (75.4%). Paying $3.18 more per task for those 1.2 points comes to $265 per task's worth. Sol at max goes further, scoring 78.0% for $1.15 and beating Astra at medium. That 78.0% is the best score from Terra or Sol. Astra beats it at three levels, high, xhigh, and max, with 78.3%, 79.0%, and 80.9%.

If the score gap is a gap in partial credit, converting it to "tasks' worth" is only a rough guide. The effort a person needs to finish partly completed work is not necessarily proportional to the number of criteria left.

In computer use, safety when the agent reads outside content also feeds into cost. The system card gives robustness to indirect prompt injection of 98.770% for Terra, 99.789% for Astra, and 98.983% for Sol. Indirect prompt injection is an attack that embeds instructions in third-party content, such as a web page, to divert an agent from the user's task. Expressed as failures, that is about 12 per 1,000 attacks for Terra and about 2 for Astra. OpenAI describes Terra, like Sol and Astra, as substantially robust to indirect prompt injection. Even so, for work that operates while reading email or outside sites, you need to budget for this gap as review and approval effort.

Under the Preparedness Framework, Terra was assessed as Critical in cybersecurity. OpenAI applies the safeguards it built for Astra to Terra as well.

System card addendum / Agents' Last Exam leaderboard

Running it yourself, screenshots and caching set the cost

The costs in OpenAI's charts were measured in OpenAI's own evaluation setup. When you call the API from your own application, the cost depends on how many screenshots you send and how often the cache hits. According to OpenAI's computer use guide, screenshots are billed as image input.

The token count of an image comes from covering it with 32 by 32 pixel patches and multiplying the patch count by a per-model multiplier. In OpenAI's table of calculation rules, every model from GPT-5.2 onward, including GPT-6 Astra, uses 1.2. As of October 27, Terra was not in that table. Here we assume Terra also uses 1.2, the same as Astra. A full HD screen (1920 by 1080) is 60 patches wide and 34 tall, 2,040 patches in all, which comes to 2,448 tokens per screenshot at 1.2.

Consider a task that takes 50 screenshots and sends the full history with every request. The nth request contains n images, so total image input is 2,448 × (1 + 2 + … + 50). The sum from 1 to 50 is 1,275, giving about 3.12 million tokens. We assume the 122,400 tokens of newly added images are billed at the cache-write rate and the remaining roughly 3 million tokens are read from cache. With 30-minute retention, cache writes cost 1.25 times the input rate.

On Terra, writes cost 122,400 tokens × $1.25 ÷ 1 million, about $0.15, and reads cost about 3 million tokens × $0.05 ÷ 1 million, about $0.15. The total is about $0.30. On Astra, writes cost about $1.53 and reads about $3.00, for a total of about $4.53. The gap is about 15 times, larger than the tenfold difference in rates. Terra's cached input is 5% of its input rate, while Astra's is 10%, so the discounts differ.

If the cache never hits, all 3.12 million tokens are billed at the normal input rate: about $3.12 on Terra and about $31.21 on Astra. On Terra, whether the cache hits alone changes image cost by about 10 times. This estimate counts images only and excludes instructions, tool definitions, and the model's output and reasoning tokens.

Long tasks have one more price threshold. When a single request has more than 272K input tokens, the whole request is billed at twice the input rate and 1.5 times the output rate. Counting images alone, a little over 110 screenshots reach that threshold. Dropping old screenshots from the history makes the threshold harder to cross. However, the cache is reused only for a matching prefix from the start. If you drop an image in the middle, nothing after that point can be read from cache. Where to drop images is a decision to make while measuring cost.

Computer use guide / Calculating image input costs / Prompt caching guide

How to choose

What follows is LATENT's suggestion. For work that repeats the same steps in volume, like entering invoices, start by trying Terra at low or medium. On Agents' Last Exam, Terra at low was both cheaper and higher scoring than Sol at low.

For work that takes many steps across desktop applications, measure Sol at medium alongside it. The OSWorld 2.1 figures show that Terra's lower rate does not always mean a lower cost per task.

Reasons to choose Astra are work where cleaning up failures is expensive and work that operates while reading outside web pages or email. Astra beats Terra on the top OSWorld 2.1 scores and on robustness to indirect prompt injection. If you compare on computer-use scores alone, OpenAI's chart also makes Claude Sonnet 5.5 a candidate.

In ChatGPT Work and Codex, cost is set by your plan's usage limits. According to OpenAI, Terra handles more tasks than GPT-5.6 Terra within the same limits. Compared with Sol and Astra, the number grows substantially. OpenAI has not published specific task counts.

If you are evaluating a deployment of your own, you need to measure four values on your own work. The first is the input, cached input, and output tokens per task. The second is the number and resolution of screenshots, and the third is the share of tasks completed end to end. The fourth is the time people spend on fixes. With all four, you can compare the total cost per task for each model.

LATENT covered the GPT-6 Astra launch on September 4 and the price contest between GPT-6.1 Sol and Claude Opus 5.5 on September 30. Terra is the high-volume option placed below those two models.

Sources and scope

Pricing, specifications, and chart values were checked against primary sources on October 27, 2026. The release date follows OpenAI's news listing. Cost per task and scores are OpenAI's published figures from its own evaluation setup, and LATENT did not measure them. The values for Claude Opus 5.5 and Claude Sonnet 5.5 are also OpenAI's measurements. The charts do not state how scores are calculated, and we could not confirm whether they are averages including partial credit or the share of tasks meeting every criterion. Terra's image token multiplier is unpublished, so the screenshot estimate assumes 1.2, the same as Astra.