Quick answer: GPT-6.1 Terra is OpenAI’s new speed-and-cost-focused model for visual reasoning, computer use, focused coding and everyday professional writing. You can use it in ChatGPT Work and Codex on eligible paid plans, through the OpenAI API as gpt-6.1-terra, in GitHub Copilot, Amazon Bedrock and Microsoft Foundry. It is not available in ordinary Chat, and ChatGPT Work does not currently support UAE inference residency for Terra.[1][2][3]
GPT-6.1 Terra setup checklist
- [ ] In ChatGPT desktop or Work on the web, open the model control, choose Advanced, then select GPT-6.1 Terra.[4]
- [ ] If it is missing, confirm that you are in Work or Codex rather than Chat, that your plan is eligible and that a workspace administrator has not disabled it.[1][4]
- [ ] For API projects, use the exact model ID
gpt-6.1-terra.[5] - [ ] Use the Responses API when you need tool calling; the model also supports Chat Completions for text and image input.[5]
- [ ] Start with the default
mediumreasoning effort, then test lower effort for speed or higher effort for difficult visual and computer-use tasks.[5] - [ ] Add a human confirmation step before payments, deletion, publishing, account changes or other consequential computer actions.
- [ ] Measure your own task accuracy, latency and full tool cost before replacing another model.
Who can access GPT-6.1 Terra?
OpenAI says Terra is available in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users; its product documentation also lists Business. Work supports Terra in the desktop app, on the web and on mobile. Enterprise access can depend on workspace availability and administrator controls.[1][4]
The important distinction is that Terra is not offered in the normal Chat experience. If you keep checking the standard Chat model picker, you may not see it even with an eligible subscription. In the desktop app or ChatGPT Work, open the model list under Advanced instead.[4]
For developers, the model is available through the OpenAI API. It is also generally available through Amazon Bedrock and Microsoft Foundry.[1][2][7]
GitHub says Terra is rolling out to Copilot Pro, Pro+, Max, Business and Enterprise users. It can be selected in Visual Studio Code, Visual Studio, Copilot CLI, the cloud agent, the Copilot app, github.com, GitHub Mobile, JetBrains IDEs, Xcode and Eclipse. Business and Enterprise administrators can control it with Copilot model policies.[3]
GPT-6.1 Terra pricing
OpenAI’s standard short-context API prices are:
- Input: $1 per 1 million tokens
- Cached input: $0.05 per 1 million tokens
- 30-minute cache write: $1.25 per 1 million tokens
- Output: $5 per 1 million tokens[5][6]
Those rates apply to requests with no more than 272,000 input tokens. For longer requests, OpenAI lists $2 per million input tokens, $0.10 per million cached input tokens and $7.50 per million output tokens. Batch and Flex are 50% below Standard, Fast mode is twice the Standard rate, and eligible regional processing adds a premium.[5][6]
Simple cost examples
Using standard short-context rates and excluding tools, storage, cache writes and regional premiums:
- 1 million input tokens + 100,000 output tokens = $1.50
- 10 million input tokens + 1 million output tokens = $15
- 100 million input tokens + 10 million output tokens = $150
Formula:
Estimated token cost = (input tokens ÷ 1,000,000 × $1)
+ (output tokens ÷ 1,000,000 × $5)
Tool charges can materially change the total. OpenAI lists separate fees for web search, file search and hosted containers, while tokens consumed by tools are billed at the selected model’s rates.[6]
Terra vs GPT-6.1 Sol: which should you choose?
Use GPT-6.1 Terra first when the job is high-volume, visually grounded and well scoped: reading forms, checking screenshots, matching catalog images, drafting from supplied material, navigating an app or making a focused code fix. OpenAI positions it as the faster, lower-cost option, with standard token prices 50% below GPT-6.1 Sol.[1][6]
Keep GPT-6.1 Sol for harder coding and reasoning where quality matters more than speed or token cost. OpenAI explicitly says Sol and GPT-6 Astra remain better choices for the most complex coding tasks.[1]
Do not choose only from benchmark headlines. Run the same representative tasks through both models, use the same reasoning effort where possible, and compare:
- Whether the final artifact is correct
- Whether computer actions landed in the intended app or file
- How often a person had to correct the model
- End-to-end latency
- Token plus tool cost
A cheaper model is not cheaper if it creates enough rework to erase the savings.
What GPT-6.1 Terra supports
The OpenAI API documentation lists:
- Text and image input with text output
- A 1,050,000-token context window
- Up to 922,000 input tokens and 128,000 output tokens
- Reasoning effort settings from
nonethroughmax - Streaming, structured outputs, function calling, file search, web search and prompt caching
- Computer use, code interpreter, hosted shell, MCP, skills and other tools through the Responses API[5]
It does not support audio or video input, Realtime, fine-tuning, embeddings, speech generation or transcription through the model’s listed endpoints.[5]
OpenAI reports that Terra improves on GPT-5.6 Terra in visual reasoning, computer use, factuality and terminal coding. These are vendor evaluations, so treat them as useful launch data rather than proof that the model will win on your own workload.[1]
OpenAI API quick start
Set your API key in an environment variable rather than placing it directly in source code. Then send a basic Responses API request:
curl https://api.openai.com/v1/responses \
-H "Authorization: Bearer $OPENAI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "gpt-6.1-terra",
"reasoning": {"effort": "medium"},
"input": "Turn these meeting notes into a five-item implementation checklist. State any missing owner or deadline instead of guessing."
}'
Tool calling requires the Responses API. Chat Completions is supported for simpler text-and-image requests, but the Responses API is the safer default if the application will later need tools.[5]
A practical model-selection prompt
Use a bounded prompt that defines the artifact, evidence and stop conditions:
Review the attached campaign screenshots for cropped text, alignment problems,
wrong destination URLs and inconsistent calls to action.
Return:
1. A table with screenshot name, issue, severity and exact fix
2. A separate “no issue found” list
3. Anything you could not verify
Do not edit, publish or upload files. Do not infer text that is unreadable.
This uses Terra’s visual strengths while keeping the first run read-only. After the review is accurate, add computer actions one controlled step at a time.
Best use cases
1. Visual quality assurance
Give Terra screenshots, product photos, floor plans or technical drawings and ask it to identify specific discrepancies. Both OpenAI and Microsoft highlight visual-to-action workflows such as checking campaign assets, interpreting drawings and updating software based on what appears on screen.[1][7]
2. Repetitive computer tasks
Terra is designed to combine screen understanding with computer actions. Good candidates include copying verified invoice fields into accounting software, updating catalog records and completing routine forms.[1][7]
Start in a sandbox or test account. Require confirmation before sending, deleting, purchasing or changing permissions.
3. Focused coding and terminal work
Use Terra for bounded fixes, test updates, repository investigation and terminal tasks where speed and cost matter. GitHub describes it as an everyday agentic coding model and says rollout is gradual across Copilot clients.[3]
4. Clear business writing
Terra can draft and revise proposals, summaries and emails from supplied information. Give it the audience, desired length, source material and facts it must not invent.[1][4]
When not to use Terra
Choose another model or workflow when:
- The task requires audio or video as direct model input.[5]
- You need Realtime voice, transcription, embeddings or fine-tuning.[5]
- The coding problem is highly complex and Sol or Astra produces materially better results.[1]
- Your ChatGPT deployment requires UAE inference residency; OpenAI says Terra in Work does not currently support it.[4]
- The action is consequential and you cannot place a person or deterministic control before execution.
Safety checklist for computer use
- Start with observation and a proposed action plan.
- Use a separate test account with minimum permissions.
- Keep credentials out of prompts and screenshots.
- Restrict websites, apps, folders and tools to the minimum needed.
- Require confirmation before send, submit, publish, delete, purchase or permission changes.
- Record the requested action and the actual saved result.
- Verify the final state in the target system rather than trusting a completion message.
- Stop when the page, data or requested target is ambiguous.
Availability outside OpenAI
Amazon Bedrock supports Terra with text and image input, a context window of up to 1 million input tokens, and Bedrock governance and audit controls. AWS says customers can call it through Bedrock APIs or configure ChatGPT Work and Codex to use it on Bedrock.[2]
Microsoft Foundry offers Terra through the Responses API in Global Standard deployments and Data Zone deployments in the US, EU and APAC. Microsoft’s published Foundry pricing varies by deployment and context length, so use the platform’s own calculator rather than assuming direct OpenAI API rates.[7]
GitHub Copilot uses provider list pricing under usage-based billing, and access can be controlled by an organization’s model policy. Because rollout is gradual, an eligible user may not see Terra immediately.[3]
FAQ
Is GPT-6.1 Terra available in regular ChatGPT Chat?
No. OpenAI’s current product note says Terra is available in ChatGPT Work and Codex but not in Chat.[4]
What is the GPT-6.1 Terra API model name?
Does GPT-6.1 Terra accept images?
Yes. It accepts text and image input and produces text output.[5]
Can GPT-6.1 Terra use tools?
Yes. OpenAI lists web search, file search, code interpreter, hosted shell, computer use, MCP and other tools. Tool calling requires the Responses API.[5]
Is GPT-6.1 Terra cheaper than GPT-6.1 Sol?
At OpenAI’s standard short-context API rates, Terra is priced at $1 per million input tokens and $5 per million output tokens, while Sol is $2 and $10 respectively.[6]
Why is GPT-6.1 Terra missing from my model picker?
Check that you are using Work or Codex rather than Chat, that the signed-in plan is eligible, and that your workspace administrator allows the model. OpenAI and GitHub also note that enterprise controls and gradual rollout can affect availability.[3][4]