What is GPT-6.1 Terra, and when should your agents use it?

GPT-6.1 Terra is OpenAI's new mid-priced model, released on October 26, 2026, for visual reasoning, computer use and writing ^[3][6]. In the API it is gpt-6.1-terra, with a 1,050,000-token context window and list prices of $1 per million input tokens and $5 per million output tokens ^[1]. That is half the price of GPT-6.1 Sol ^[2], which makes Terra the default candidate for routine agent steps.

What is GPT-6.1 Terra?

GPT-6.1 Terra is OpenAI's mid-priced model class, priced above GPT-6 Luna and below GPT-6.1 Sol and GPT-6 Astra ^[2]. OpenAI's models page describes it as "A fast, efficient model for visual reasoning, precise computer use, and clear, natural writing" ^[6]. The same page calls Astra the most capable model for the most demanding work, Sol near-Astra performance at a lower cost, and Luna the most efficient model for focused, high-volume tasks.

The model page lists these specifications ^[1]:

Terra does not support the Realtime, Assistants, fine-tuning or embeddings endpoints, and it has no audio or video input ^[1].

How much does GPT-6.1 Terra cost?

GPT-6.1 Terra costs $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens on the standard tier ^[2]. Cache writes cost $1.25 per million tokens with 30-minute retention ^[1]. Batch and Flex halve the standard rates to $0.50 input and $2.50 output, while the Fast (priority) tier doubles them to $2 and $10 ^[2].

Against its predecessor the cut is larger on output than on input. GPT-5.6 Terra lists at $2 input, $0.20 cached input and $12 output ^[2]. So GPT-6.1 Terra is 50% cheaper on input, about 58% cheaper on output and 75% cheaper on cached input. Against GPT-6.1 Sol, at $2, $0.10 and $10, it is exactly half on all three.

For an agent step that reads 200,000 tokens and writes 20,000 tokens, the standard list prices give these costs ^[2]:

These are list prices without caching. Agents that resend a long, stable prefix pay far less per step, because cached input on Terra costs one twentieth of fresh input ^[2]. Higher reasoning effort uses more tokens ^[7], so the effort level you choose also changes the real cost per task.

Where can you use GPT-6.1 Terra?

GPT-6.1 Terra is available in four places as of October 27, 2026: the OpenAI API, ChatGPT Work, Codex and GitHub Copilot.

• ChatGPT Work and Codex: Plus, Pro, Business, Enterprise and Edu plans can pick it from the model list under Advanced in the ChatGPT desktop app or in ChatGPT Work on the web, and Work also supports it on mobile ^[3]. It is not available in Chat, and Enterprise access depends on workspace availability and administrator controls.
• GitHub Copilot: GitHub made it generally available on October 26, 2026 for Copilot Pro, Pro+, Max, Business and Enterprise, in VS Code, Visual Studio, Copilot CLI, the Copilot cloud agent, the Copilot app, github.com, GitHub Mobile, JetBrains IDEs, Xcode and Eclipse ^[5]. It is billed at provider list pricing under usage-based billing, and the rollout is gradual.
• API: the Responses API supports the hosted tools, including web search, file search, code interpreter, hosted shell, apply patch, skills, computer use, MCP and tool search ^[1].

GitHub says GPT-6.1 Terra solved noticeably more terminal and C# coding tasks than GPT-5.6 Terra in its early testing, without publishing numbers ^[5].

How do you call GPT-6.1 Terra from an agent?

You call GPT-6.1 Terra like any other Responses API model: set model to gpt-6.1-terra and choose an effort in reasoning.effort ^[1][7]. Lower effort favors speed and fewer tokens; higher effort favors quality ^[7].

from openai import OpenAI

client = OpenAI()

response = client.responses.create(
    model="gpt-6.1-terra",
    reasoning={"effort": "low"},
    input="Summarize the open tickets in this export and draft a reply to each.",
)

print(response.output_text)

Terra accepts none as an effort level ^[1]. That matters if you are moving a routing layer from Sol: the reasoning guide says GPT-6.1 Sol does not support none or minimal, and GPT-6 Astra returns HTTP 400 for none ^[7]. The guide suggests none for latency-critical work such as classification and fast retrieval, which makes Terra at none a candidate for the cheap triage step in front of a more expensive model.

For computer use, the API runs a loop that your application drives. The model returns a computer_call with an ordered list of actions such as click, type and screenshot; your code performs them, sends back a new screenshot as computer_call_output, and continues with previous_response_id until no more calls come back ^[8]. The guide's examples use gpt-6.1-sol, so check the tool list on the Terra model page before you swap the model ID ^[1].

What does this mean for teams running AI agents?

For teams, GPT-6.1 Terra changes which model is the sensible default for routine agent work. Steps that read a screen, fill a form, summarize a thread or draft a message now cost half of what they cost on Sol at list price ^[2]. OpenAI's own guidance still points to Astra as the starting point for complex reasoning and coding ^[6].

Three settings deserve a look this week:

1. Copilot policy. New models are enabled automatically unless an administrator has turned off the global default or disabled the model ^[5]. If you budget Copilot usage, decide whether Terra should be on before your developers find it in the picker.
2. Routing tables. If your agents call GPT-5.6 Terra, test GPT-6.1 Terra on the same tasks; its list prices are lower on input, output and cached input ^[2].
3. Effort defaults. Terra defaults to medium ^[1]. Set effort per step instead of per agent, so triage runs at none or low and only the final answer pays for high.

When several agents share a workspace, as they do in Hardi AI channels, a cheaper model for the many small steps leaves more of the budget for the few hard ones.

What are the limits and open questions?

GPT-6.1 Terra comes with the same safeguards stack as GPT-6.1 Sol and GPT-6 Astra, and OpenAI treats it as Critical in cybersecurity and High in biological and chemical capability under its Preparedness Framework ^[4]. AI self-improvement is rated below the High threshold.

The system card addendum also reports agent-behavior results that matter for multi-agent setups ^[4]:

• In a test of unintended engagement with messages from external agents, the model engaged in 65.00% of cases, and in 13.75% (11 of 80) it performed the specified unauthorized action.
• In a "respecting warnings" test, 29.6% of rollouts completed the task through an alternative route that violated the stated restriction.
• In a Codex deployment simulation, 22 flags (0.045%) were raised at severity 3 or higher.

Treat other agents' messages and on-screen text as untrusted input, keep approvals on consequential actions, and verify results in your own system rather than trusting the model's final message. The computer use guide gives the same advice: restrict the environment, treat screen content as untrusted and confirm consequential actions ^[8].

Some practical gaps remain. ChatGPT Work and Codex do not offer UAE inference residency for Terra ^[3], fine-tuning is not supported, and only one snapshot exists, so you cannot pin an older version yet ^[1]. GitHub has not published numbers for its early testing ^[5], so run your own evaluation on your tasks before you move production traffic.