OpenAI released GPT-6.1 Terra through the Responses and Chat Completions APIs on October 26, adding a lower-priced model aimed at workloads that include computer use. The standard API rate is $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens.
Terra is also rolling out in ChatGPT Work and Codex, but it is not available in the standard ChatGPT experience. That distinction matters for businesses communicating access to customers: an API listing, a workplace product rollout and availability in the consumer chat interface are separate distribution channels.
The launch creates a new OpenAI routing option between inexpensive models used for classification or extraction and more costly reasoning models. It also gives gateway operators a familiar operational problem: the headline price is only the beginning of the integration work.
Terra’s price depends on workload shape
At the published rates, output tokens cost five times as much as ordinary input tokens. Cached input is priced at 5% of the standard input rate. Applications that repeatedly send stable instructions, tool descriptions or other reusable context could therefore see a materially different cost profile from applications that generate large responses or continuously change their prompts.
This is especially relevant to computer-use agents. Such systems can run through multiple observation, reasoning and action cycles before completing a task. Even when the model’s input price is low, verbose outputs or long tool loops can raise the total bill. OpenAI’s announced rate should therefore be evaluated against complete task cost, not only the price of the first request.
Teams using unified AI API billing will need separate accounting for ordinary input, cached input and output. Collapsing those categories into one blended token figure would obscure both customer charges and gateway margins. Evaluations should record tokens and cost across the full agent run, including failed steps and retries.
Gateways need metadata, policy and compatibility updates
Adding Terra to a multi-model API gateway for teams requires more than registering its model ID. Catalog records need the three token rates, supported API surfaces, capability metadata and regional restrictions. Routing systems also need to distinguish whether a workload benefits from prompt caching and whether it is likely to produce enough output or tool iterations to erase the apparent input-price advantage.
OpenAI says UAE inference residency is not supported for this model. A gateway that promises regional handling cannot treat Terra as an interchangeable substitute for another OpenAI model when a tenant requires UAE residency. The restriction should be visible in model discovery, enforced in routing policy and preserved in request logs so operators can explain why a request was blocked or sent elsewhere.
Resellers face an additional communication issue. An AI API reseller platform should not imply that a customer’s ChatGPT subscription includes Terra API access, or that API availability means the model can be selected in standard ChatGPT. Product catalogs and customer-facing usage reports should identify the access channel and effective model separately.
For Model Gate and similar OpenAI-compatible services, the practical release checklist includes adding the model, applying cache-aware prices, exposing the residency limitation and testing both Responses and Chat Completions behavior. Margin rules may also need revision if cached tokens are passed through at provider cost while ordinary input and output carry different markups.
Performance claims still need independent testing
OpenAI has reported capability, speed and hallucination improvements for Terra, but those claims have not all been independently reproduced. The right comparison will also vary by workload. A model that is economical for short structured responses may be less attractive for an agent that emits extensive reasoning, retries actions or carries out long computer-use sessions.
Developers should benchmark Terra with representative prompts, tools and completion limits rather than transferring results from conventional chat tests. Useful measurements include task completion rate, end-to-end latency, cache-hit frequency, output volume, number of tool cycles and total cost per successful task.
Terra’s immediate significance is therefore not that it is universally cheaper. It gives API buyers another OpenAI price band and gives routing systems another set of trade-offs to model. Whether it reduces production costs will depend on caching behavior, response length, agent-loop efficiency and the regional policies attached to each request.