# GPT-6.1 Terra: Pricing, API Details and Computer-Use Benchmarks

OpenAI's cheaper visual reasoning model makes everyday agent work more affordable, but the headline token price is only part of the calculation.

By Paras Madan | 2026-10-27


OpenAI released GPT-6.1 Terra on October 26, combining visual reasoning, computer use and writing at half GPT-6.1 Sol's standard API token prices. The practical question is whether a cheaper model can finish the same workflow without creating more retries or review work.

The [launch announcement](https://openai.com/index/introducing-gpt-6-1-terra/) lists availability in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, alongside the API. Terra succeeds GPT-5.6 Terra. This is a model release, not a promise that every application can now run unattended.

## GPT-6.1 Terra API pricing

The [API pricing page](https://developers.openai.com/api/docs/pricing) lists these standard rates per million tokens for short-context requests:

| Model | Input | Cached input | Output |
| --- | --- | --- | --- |
| GPT-6.1 Terra | $1 | $0.05 | $5 |
| GPT-6.1 Sol | $2 | $0.10 | $10 |
| GPT-6 Astra | $10 | $1 | $50 |

Cache writes are separate: Terra costs $1.25 per million tokens for 30-minute retention or $2.50 for 12-hour retention. A cache hit and creating the cache are different billing events, so the five-cent rate should not be treated as the cost of every repeated prompt.

For a simple illustration, 100,000 uncached input tokens and 10,000 billed output tokens cost $0.15 with Terra, versus $0.30 with Sol at standard rates. That is token arithmetic, not an estimate for a complete agent task. Tools, execution infrastructure and extra model calls can change the total.

**Measure cost per successfully completed workflow, not just cost per token.** A cheaper call that needs several attempts can erase part of the saving. Conversely, a model that completes the task with fewer steps can save more than its rate card suggests.

## Context window, reasoning and API limits

The [Terra model documentation](https://developers.openai.com/api/docs/models/gpt-6.1-terra) specifies a 1,050,000-token context window and a maximum output of 128,000 tokens. The API identifier is `gpt-6.1-terra`.

Reasoning effort supports `none`, `low`, `medium`, `high`, `xhigh` and `max`, with `medium` as the default. Tool calling requires the Responses API. Structured outputs and streaming are supported; fine-tuning is not.

There is an important pricing boundary inside that large context window. Prompts above 272,000 input tokens incur twice the input and cache rates and 1.5 times the output rate for the full request. Batch and Flex are half standard pricing, while Fast mode doubles it. Regional processing adds a premium where available.

Terra accepts text and images and produces text. Native audio and video are not supported. Its supported image-generation tool is distinct from native image output. Similarly, operating a video-editing application through its interface does not mean the API accepts video as a native input modality.

The million-token capacity is useful headroom, but it is not an argument for sending everything. For a recurring document workflow, compare a full-history approach against a smaller, task-specific input before deciding what belongs in every request.

## What changed from GPT-5.6 Terra

OpenAI reports a 15-percentage-point improvement on Agents' Last Exam at medium reasoning effort, at roughly one-third the cost of GPT-5.6 Terra. On Terminal-Bench 4.0 at maximum effort, the reported score rises from 21.5% to 50.3%.

Those are vendor-reported evaluation results, not independent measurements of a production application. OpenAI still recommends Sol or Astra for the hardest coding work. Terra's position is therefore easier to understand as a lower-cost option to evaluate for frequent tasks, rather than an automatic replacement for every larger model.

A useful migration test would hold the application, instructions and success criteria constant, then compare Terra and the current model on the same saved tasks. Include failures that matter commercially: the wrong record edited, an incomplete form submitted, or a plausible answer that cannot be traced to the source document.

## How computer use actually works

The [computer-use guide](https://developers.openai.com/api/docs/guides/tools-computer-use) describes a loop: send the task, execute the model's requested actions in a browser or desktop environment, and return screenshots so it can inspect the result. Your application still needs the execution environment and action handlers.

This distinction matters when budgeting an integration. Buying access to the model does not provide all the permissions, business rules or recovery logic around it. A browser agent entering supplier information, for example, needs a reliable way to distinguish a saved draft from a submitted record.

OpenAI's guidance calls for isolated environments, restricted sites and actions, confirmation for consequential operations, and limits on steps, time or cost. Screen content must be treated as untrusted, and the outcome should be checked in the application rather than accepted from the model's final message.

## Where Terra is worth testing first

Start with a narrow, repeatable workflow where success is observable and mistakes are reversible. Document review followed by a proposed data entry is a more controlled starting point than allowing an agent to make changes across several live business systems.

The opportunity is to make more useful work affordable without making supervision more expensive. If Terra passes the same checks as the model it replaces, the saving becomes operationally meaningful. If it needs substantially more intervention, keep the stronger model for that part of the workflow and measure again after the next improvement.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help AI founders turn technical launches and product thinking into clear stories developers and founders want to read and share.

[Book a 30-minute call](https://cal.com/paras-madan/30min).