GPT-6.1 Terra is OpenAI’s new model for visual reasoning and computer-based work, available in ChatGPT Work, Codex and the API. Its standard API rates are US$1 per million input tokens and US$5 per million output tokens. The useful question is whether it completes your workflow reliably—not simply whether its tokens are cheaper.
The official announcement positions Terra for tasks involving screens, documents and everyday writing. OpenAI lists access for Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex. For the most complex coding problems, it still points to GPT-6.1 Sol and GPT-6 Astra. Our Sol guide covers that different model choice.
What the launch video shows—and what it does not
OpenAI’s “GPT-6.1 Terra is now in orbit” video introduces the model visually, including a computer-work example that turns a tram reference into a 3D model. It is a useful illustration of the direction: understand something on screen, then work through software to produce an artifact.
A short launch demo is not a reproducible test of your business process. The written model reference and system card provide the technical capabilities and limitations discussed below.
For a real project, I would separate three checks: did the model read the reference correctly, did it operate the application correctly, and does the saved result match the request? An attractive preview can pass the first impression while still missing a dimension, a field or a required export.
API capabilities: images in, tools for action
The official model reference identifies the model as gpt-6.1-terra. It accepts text and images and produces text. Native audio and video are not supported model modalities.
That distinction matters for the video-editing examples: an agent operating an editor through tools is not the same as sending a video to a native video endpoint. Your application still needs the appropriate tools, files and execution environment.
The reference lists Responses, Chat Completions and Batch support, but tool calling requires the Responses API. Supported Responses tools include computer use, MCP and code-oriented tools. A supported tool is not an unrestricted permission grant: your integration must still control accounts, destinations and allowed actions.
The documented context window is 1,050,000 tokens, with maximum input of 922,000 and maximum output of 128,000. Those limits are not a recommendation to send every available file. Smaller, relevant context is easier to inspect and can avoid unnecessary cost. Medium reasoning effort is the default; evaluate the effort setting against the actual task.
GPT-6.1 Terra API pricing
The current pricing reference lists these standard US-dollar rates per million tokens:
| Token category | Standard price |
|---|---|
| Input | US$1.00 |
| Cached input | US$0.05 |
| Output | US$5.00 |
For illustration, 100,000 uncached input tokens plus 10,000 output tokens cost US$0.15 at those standard short-context rates, before tool fees or other modifiers. This is a token calculation, not a quoted price for completing a business task.
Cache writes are billed separately: US$1.25 per million tokens for a 30-minute duration or US$2.50 for 12 hours. Above 272,000 input tokens, the entire request uses twice the input/cache rates and 1.5 times the output rate. Tool charges and other processing choices can add to the total.
Use absolute rates when comparing models. A broad “half-price” slogan can hide differences among output rates, caching and request sizes. Measure the bill for a successfully completed task, including retries and human corrections, rather than multiplying a headline price by an optimistic token estimate.
Benchmarks do not remove the review step
OpenAI reports improvements in visual reasoning, computer use, factuality and coding. Those are publisher-run evaluations, not FindMilan benchmarks. They do not establish the same result on your software, prompts or data.
The GPT-6.1 Terra system card, published October 26, explains why evaluation settings matter: prompts, tools and reasoning effort can differ from production use. Some factuality tests deliberately concentrate on error-prone examples; their percentages should not be presented as everyday error rates.
The card also reports mixed safety outcomes. Its Critical cybersecurity capability classification is not a notice that a customer system was breached. Improved instruction following and resistance to malicious prompts do not make sandboxing, permissions or human approval optional.
My practical rule is simple: give the agent a bounded task and an observable finish condition. Keep approval before payments, publication, external messages and destructive changes. Verify the destination result instead of accepting the agent’s summary as proof.
A low-risk pilot before switching production work
Start with one repetitive workflow that has a clear answer and a safe test environment. For example, read a synthetic invoice image and prepare a draft record without submitting it to accounting software.
- Define the fields, allowed destinations and actions the agent must not take.
- Use test credentials and representative examples, including ambiguous or incomplete inputs.
- Compare Terra with your existing model using the same tools and acceptance criteria.
- Record completion quality, elapsed time, corrections and total billed usage.
- Add a human checkpoint before any consequential write, then verify the final saved state.
These are my evaluation recommendations, not a claim that I benchmarked Terra or migrated your system. Keep a stronger-model fallback for difficult cases only when your tests justify it, and make failure visible rather than silently retrying an external action.
For help choosing a model and building the surrounding controls, see AI consulting, workflow automation and the AI Web Awards project. The goal is a useful, verifiable workflow—not more autonomy than the task needs.
Frequently asked questions
Where is GPT-6.1 Terra available?
OpenAI lists Plus, Pro, Enterprise and Edu access in ChatGPT Work and Codex, plus API access as gpt-6.1-terra. Check the actual surface and account rather than assuming free-tier availability.
What does GPT-6.1 Terra cost through the API?
Standard rates are US$1 input, US$0.05 cached input and US$5 output per million tokens. Cache writes, long context, processing options and tools can change the bill.
Can the GPT-6.1 Terra API accept video or audio directly?
No native audio or video is listed. It accepts text and images; using an editor through computer tools is a separate workflow.
Which API should I use for Terra tool calling?
Use Responses for tools. Chat Completions and Batch are also supported endpoints, but neither a model selection nor tool support authorizes a consequential action.
Frequently asked questions
Where is GPT-6.1 Terra available?
OpenAI lists Plus, Pro, Enterprise and Edu access in ChatGPT Work and Codex, alongside API access as gpt-6.1-terra. This is not a promise of access on every ChatGPT surface or the free tier.
What does GPT-6.1 Terra cost through the API?
Standard rates are US$1 per million input tokens, US$0.05 per million cached input tokens and US$5 per million output tokens. Cache writes, long-context multipliers, processing options and tool fees can change the total.
Can the GPT-6.1 Terra API accept video or audio directly?
The model reference lists text and image input with text output, not native audio or video. Working in a video editor through computer-use tools is a different capability.
Which API should I use for Terra tool calling?
Use the Responses API for tool calling. The model also supports Chat Completions and Batch, but tool availability and permission to perform an action are separate concerns.