GPT-6.1 Terra: Computer Use, Workflows and API Pricing

GPT-6.1 Terra launched on 26 October 2026 as a faster, lower-cost OpenAI model for visual reasoning, precise computer use and clear writing. Developers can use it through the API as gpt-6.1-terra; eligible users can select it in ChatGPT Work and Codex. Standard API pricing is $1 per million input tokens and $5 per million output tokens, with separate cached-input pricing. It is not being offered as a normal ChatGPT Chat model.

Editorial verification: 29 October 2026. Primary sources: OpenAI's GPT-6.1 Terra launch and OpenAI's dated model release notes. Benchmark and relative efficiency figures here reflect the company's published tests, not independent RKWS measurements.

What is GPT-6.1 Terra?

GPT-6.1 Terra is a model designed for many of the visual and computer-based activities that fill a normal workday: examining screenshots, interpreting technical drawings, preparing content, using software tools and following multi-step instructions.

OpenAI positions Terra as a cost-efficient alternative to its more powerful Sol and Astra models. The aim is to let a business run more tasks within a budget while retaining useful computer-use capability. This is not the same as saying Terra is the smartest choice for every high-stakes problem or the lowest-cost model in all possible scenarios.

Where is GPT-6.1 Terra available?

OpenAI announced rollout in ChatGPT Work, Codex and the OpenAI API. Its model release notes explicitly distinguish Work and Codex from ordinary ChatGPT chat. Eligible Plus, Pro, Business, Enterprise and Edu experiences depend on the rollout schedule, permitted features, organization controls and geographic availability.

OpenAI notes that Terra does not support UAE inference residency at launch, an inference-region distinction rather than a statement that everyone in the UAE is barred from using the model. For enterprise data policies, read the current documentation for your account.

What is GPT-6.1 Terra's standard API pricing?

The published launch rate is $1 for one million input tokens, $0.05 for one million eligible cached input tokens and $5 for one million output tokens. API consumption is billed separately from any ChatGPT subscription.

OpenAI says Terra's standard API prices are 50% below those of GPT-5.6 Terra and GPT-6.1 Sol. This comparison describes specific price categories at launch, not a promise that every Terra request costs 50% less overall. Input length, output length, repeated tool calls and task success also influence total cost.

Use the official model page and live rate card before quoting a price to a client. Do not infer API charges from a user's ChatGPT Work subscription entitlement.

How good is Terra at computer use and visual tasks?

OpenAI highlighted evaluations involving software applications, visual reasoning and multi-step professional work. Its Agents' Last Exam evaluation included tasks such as turning technical drawings into a 3D model and making a time-coded inspection of video material. At medium effort, OpenAI reports Terra exceeded GPT-5.6 Terra by 15 percentage points on this suite at about one-third the cost per task.

On another computer-use measure, OSWorld 2.1, OpenAI says Terra at medium effort exceeded its predecessor's best score. On MMMU-Pro, a visual reasoning benchmark, Terra also improved over GPT-5.6 Terra across the tested settings.

These statements apply to the company's tested conditions. They do not prove that a newly connected accounting or CRM workflow will work reliably without tailored instructions, appropriate permissions and quality checks.

What do the coding benchmarks tell developers?

On Terminal-Bench 4.0, OpenAI reports a 50.3% task-solving score for Terra at maximum effort compared with 21.5% for GPT-5.6 Terra. The benchmark covers command-line software tasks such as repairing code and diagnosing application behavior.

Be careful with evaluation comparisons: particular model settings and test configurations matter. OpenAI explicitly recommends GPT-6.1 Sol or GPT-6 Astra for its hardest coding tasks. A higher benchmark score should guide a test plan, not replace an evaluation on a real repository.

Read our AI coding agents guide for how tests, reviews and ownership fit into AI-assisted software engineering.

What does computer use mean in a real business workflow?

Computer use means a model interprets an interface and helps carry out actions with supported tools. A workflow could involve opening an authorized system, reading a permitted field, choosing the next step and entering a value. The model supplies decisions and instructions; the surrounding environment controls access and whether actions are allowed.

For example, a business receiving invoices might want a system to extract invoice data, check it against expected records and prepare a draft entry in accounting software. That workflow needs document validation, duplicate checking, authorized credentials and a human approval step before financial records are finalized.

1. Define the desired outcome and acceptable error rate.
2. Provide the agent only the relevant screens, files and accounts.
3. Restrict actions to safe data entry or draft preparation initially.
4. Require confirmation for sending messages, payments and destructive changes.
5. Log the steps and compare final records against source documents.
6. Measure completion time, errors, corrections and cost per accepted task.

When should a team choose Sol or Astra instead of Terra?

Use a model-selection strategy based on the task's complexity, risk and measured output quality. Terra is intended to be efficient for regular work; stronger models may be justified for harder reasoning or research.

There is no universal winner. Use an offline evaluation with representative jobs and compare total accepted outcomes per unit of cost. Our OpenAI Decisions API guide describes a more specialized alternative for bounded classification and routing tasks.

What safety concerns come with more capable computer use?

OpenAI's Terra system-card addendum describes the model as meeting a Critical threshold in cybersecurity capability under its Preparedness Framework. The company says it uses safeguards associated with GPT-6 Astra. That rating indicates capability and risk evaluation; it is not a certification that a customer deployment is secure by default.

Models using computers can encounter malicious instructions hidden in webpages, incorrect screen labels or stale records. Preventing unauthorized actions requires controls outside the model.

• Use dedicated test or sandbox environments wherever possible
• Grant the minimum necessary access
• Validate outputs against business rules before committing changes
• Require explicit approval for consequential actions
• Retain audit logs within the organization's data policy
• Prepare a fallback when the model is unavailable or uncertain

Our AI agent prompt-injection guide and governance checklist explain the underlying principles.

How can an Indian software company evaluate Terra?

Start with a task that is frequent, reversible and easy to score. Choose something like extracting fields from a company-approved template or preparing a content draft, not posting irreversible changes to live business systems.

1. Gather 30–50 representative examples with expected outcomes.
2. Run the task with the current method and record review time.
3. Test Terra using the same data and permitted tools.
4. Measure factual accuracy, failed steps, latency and overall API consumption.
5. Escalate ambiguous or policy-sensitive cases for human review.
6. Expand only if the accepted result and operational effort improve.

Our AI agent observability guide offers additional metrics for monitoring recurring workflows after deployment.

Frequently asked questions about GPT-6.1 Terra

Is GPT-6.1 Terra available in normal ChatGPT chat?

No. The October model release notes describe Work and Codex availability, not standard Chat.

What is the Terra API model ID?

gpt-6.1-terra, as shown in the OpenAI API release documentation.

What does Terra cost per million tokens?

At launch: $1 input, $0.05 eligible cached input and $5 output, with API usage billed separately from subscriptions.

Can it perform computer-use tasks?

It is designed for visual reasoning and computer use with supporting tools, but permissions, confirmation and validation remain necessary.

Should I choose Terra over Sol or Astra?

Choose according to tests on your real workload. OpenAI positions Sol and Astra for particularly difficult coding and research tasks.

Conclusion: efficiency does not replace verification

GPT-6.1 Terra creates a useful new price-performance option for software and business workflows. Its value depends on making more tasks accurate and affordable, not on granting unrestricted control to a model. Confirm access, monitor token costs and require review where necessary. To assess a business automation workflow, contact RKWS.

Official sources: Introducing GPT-6.1 Terra, 26 October 2026; OpenAI Model Release Notes; Terra safety addendum.

Frequently asked questions