GPT-6.1 Terra is an API model for visual work, application interaction and writing, with a lower unit price than GPT-6.1 Sol. It is not a new name for the older GPT-5.6 Terra. If you are wiring a screenshot-driven agent into an application, that distinction matters more than the family name.
OpenAI’s API changelog dates the release to October 26, 2026. This guide covers the direct API and distinguishes it from gateway labels and Copilot access. It is documentation-based guidance, not a hands-on benchmark or a claim that Terra wins every coding task.
Specifications that affect integration
The official model reference identifies gpt-6.1-terra, a 1,050,000-token context window, 128,000-token maximum output and June 1, 2026 knowledge cutoff. Text and images are inputs; output is text. Audio and video are not supported modalities on that reference.
Do not interpret the context number as a guaranteed usable repository size. Instructions, conversation history, images, tool results and output need room. A repository index can be more useful than dumping every generated file into the prompt. See AI application architecture for the surrounding request and state design.
Terra offers reasoning efforts from none through max, with medium as the documented default. Tool calling requires Responses. A plain text request and an agent executing file or browser actions therefore have different integration requirements, even if both select the same model name.
Direct API pricing
These are US-dollar prices per million tokens from OpenAI’s pricing page, checked October 29. They are token rates, not subscription prices or completed-task quotes.
| Processing | Input | Cached input | 30-minute cache write | Output |
|---|---|---|---|---|
| Standard, up to 272K input | $1.00 | $0.05 | $1.25 | $5.00 |
| Batch / Flex, same context tier | $0.50 | $0.025 | $0.625 | $2.50 |
| Standard, above 272K input | $2.00 | $0.10 | $2.50 | $7.50 |
The long-context adjustment applies to the full request, not merely the tokens beyond the threshold. A request just over that boundary can therefore change cost materially. Cache retention also matters: the published Standard 12-hour write rate is $2.50 before long-context adjustments. Fast processing costs twice Standard; regional processing has its own documented uplift where available.
For budgeting, keep input, cache writes, cache hits, output and tool charges separate. As simple arithmetic, one million uncached input tokens plus 100,000 output tokens costs $1.50 at short-context Standard rates, before tools and other charges. That is an illustrative bill calculation, not a measured workload result. Our API pricing comparison covers broader provider selection.
API access without confusing product entitlements
Use an API project with appropriate access, billing and limits. Availability in a consumer subscription or in GitHub Copilot is not evidence that the same account has direct API quota. Keep provider credentials server-side and apply environment and secret controls.
This minimal request illustrates the model selection. It was not executed against a paid endpoint for this guide:
from openai import OpenAI
client = OpenAI()
response = client.responses.create(
model="gpt-6.1-terra",
reasoning={"effort": "medium"},
input="Review this proposed UI change for missing error states.",
max_output_tokens=1500,
)
print(response.output_text)The SDK reads the normal OpenAI API credential from the environment. Do not paste a key into frontend code or a committed example. The output cap constrains generation, but it is not a substitute for application-level request and spending limits.
For tool workflows, add a validated tool contract, explicit approvals and a bounded execution loop. The model producing an argument is not authorization to run it. Start with read-only tools, inspect traces, and require human approval before destructive actions. Our agent security guide explains that separation.
Terra Pro: mode and route, not an invented direct endpoint
OpenAI’s reasoning documentation separates reasoning.mode from reasoning.effort. Pro mode asks for more model work; effort controls reasoning within the mode. Pro aggregates and bills the work performed at the selected model’s token rates. Equal unit pricing does not imply an equal final bill.
For an eligible Responses request, the relevant configuration is the existing model plus a reasoning object such as {"mode":"pro","effort":"medium"}. Do not invent gpt-6.1-terra-pro as a direct OpenAI endpoint just because another provider uses that label.
OpenRouter’s Terra Pro route is a separate access surface. Its catalog exposes openai/gpt-6.1-terra, openai/gpt-6.1-terra-pro and a :batch variant of each. Current short-context Standard and Pro routes list $1 input and $5 output; batch routes list $0.50 and $2.50. Gateway access, supported parameters and routing restrictions remain gateway-specific.
This distinction belongs inside the Terra guide because it changes how you select execution, not the identity of a separately documented direct model. If a future release introduces a genuinely separate model, capability contract or independent pricing ladder, it would warrant a new ownership review.
Coding and visual-agent use cases
Terra is a sensible evaluation candidate when the task combines screenshots with constrained changes: checking a form’s error states, mapping a product image to catalog fields, interpreting a drawing, or proposing an interface patch. Those are candidate workflows, not evidence of reliable coordinate selection or correct code on your project.
For coding, evaluate the patch rather than the explanation. Use the same repository snapshot, dependency versions, test fixtures and tool permissions when comparing Terra with Sol or another model. Record failed runs and retries. A cheaper token can become an expensive result when an agent repeatedly edits the wrong component.
For computer use, test misleading page content, ambiguous buttons and approval boundaries. Keep credentials and payments outside an unrestricted loop. A visual model can misunderstand the current screen even when its written description sounds confident. Connect this work to testing AI-generated code and LLM regression testing.
My take
Evaluate Terra as a lower-cost visual and application-workflow model, not as a universal replacement for your strongest coding model. Start with a narrow suite where screenshots or document images actually matter. Keep Pro mode off the default path until the extra work improves accepted outcomes enough to justify its latency and bill.
FAQ
When did GPT-6.1 Terra launch?
OpenAI’s API changelog records October 26, 2026. The publication date of this guide is not the model’s release date.
Does a million-token context remove the need for retrieval?
No. Select relevant files and reserve output space. Context capacity is not a guarantee of equally accurate attention to every token.
Can I use Terra tools through Chat Completions?
The official model reference requires Responses for tool calling. Do not assume that a compatible text endpoint provides the same tool contract.
Is Terra Pro a separate direct OpenAI model ID?
The current direct documentation supports Pro reasoning mode on the model. OpenRouter’s Pro route is a gateway label, not proof of a separate direct endpoint.
Have you independently benchmarked Terra?
No. The specs and prices above are documented facts; the evaluation and deployment recommendations are editorial guidance.