OpenAI launched GPT-6.1 Terra on 26 October 2026, introducing a faster, lower-cost model focused on visual reasoning, computer use, coding and everyday professional tasks. Standard API pricing starts at $1 per million input tokens and $5 per million output tokens, half the rates of GPT-5.6 Terra and GPT-6.1 Sol.
OpenAI says Terra completes certain computer-use benchmark tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol, while costing roughly nine times less per task than Claude Opus 5.5 in a specific evaluation. Those comparisons are promising, but they don’t guarantee universal speed or price. A more useful question for developers is whether Terra can reliably finish routine, multi-step work without escalating to a more expensive model.
GPT-6.1 Terra pricing and availability at a glance
| Specification | GPT-6.1 Terra |
|---|---|
| Standard API input | $1 per million tokens |
| Standard API output | $5 per million tokens |
| Cached input | $0.05 per million tokens |
| Context window | 1,050,000 tokens |
| Maximum output | 128,000 tokens |
| API model ID | gpt-6.1-terra |
| Access | ChatGPT Work, Codex and OpenAI API |
The prices above are standard text-token rates, not an all-inclusive charge for an agent that searches the web, controls a computer or calls other tools. OpenAI’s pricing rules also change for very large prompts, so don’t apply the headline $1 rate to every workload without checking its size.
Computer use is the headline improvement, not just lower token prices
Terra is designed to interpret what appears on a screen and take subsequent actions. OpenAI highlights tasks such as reading technical drawings, checking visual assets, completing forms and working across applications. This makes the release particularly relevant to teams evaluating AI agents rather than simply looking for a cheaper chatbot.
On Agents’ Last Exam, a benchmark covering professional software-based tasks, OpenAI reports that GPT-6.1 Terra scores 15 percentage points higher than GPT-5.6 Terra at medium reasoning effort, at approximately one-third of the cost. It also says Terra performs on par with GPT-6 Astra on that evaluation while carrying one-tenth of Astra’s standard per-token price.
On the same computer-use comparison, OpenAI reports task completion more than three times faster than Astra and GPT-6.1 Sol. A faster model can reduce waiting time across a lengthy agent workflow, but only if it avoids repeated attempts, misread screens and unnecessary tool calls. Those behaviours need testing in the application where the model will actually run.
Why the nine-times-cheaper claim needs context
The comparison with Claude Opus 5.5 uses maximum reasoning effort on Agents’ Last Exam. OpenAI says Terra uses more than three times fewer tokens for this work and costs about one-quarter of Opus 5.5 per token, resulting in about nine times the lower cost per task on that benchmark.
That does not mean a coding session, research report, or customer-service workflow will automatically cost nine times less. The result depends on the task mix, reasoning setting, response length, retries, and any paid tools involved. Nor does similar performance to Astra on one benchmark establish equal capability across difficult reasoning tasks.
Developers comparing models should measure the cost of an accepted result, including failed attempts and human corrections, rather than rank them by input-token price. DIY AI’s AI model pricing comparison explains how those extra costs can overturn an apparently cheaper choice.
Coding and factual accuracy improve, with important limits
OpenAI reports a Terminal-Bench 4.0 score of 50.3% for GPT-6.1 Terra at maximum reasoning effort, up from 21.5% for GPT-5.6 Terra. This is a substantial improvement on the published benchmark, although OpenAI still recommends GPT-6.1 Sol or GPT-6 Astra for the most complicated software engineering tasks.
In its internal test of prompts designed to induce factual errors, the company reports that error rates at low reasoning effort fell from 53.6% to 16.7%. At maximum effort, it reports an error rate of 4.3%. These are results from OpenAI’s evaluation set, not measurements of the probability that any answer to a real user will be wrong.
For developers, the sensible test is whether Terra preserves requirements over several edits, catches its own mistakes and passes an existing test suite. A lower benchmark error rate does not remove the need to verify code changes, financial entries or actions taken in a live account.
The $1 API rate has exceptions for longer workflows
A straightforward request with 20,000 uncached input tokens and 5,000 output tokens would cost about $0.045 at standard rates, before any tool charges. This illustrates why Terra could be economical for repetitive document and screen-based tasks.
However, OpenAI’s API documentation specifies that requests exceeding 272,000 input tokens incur double the standard input and cache rates and 1.5 times the standard output rate, applied to the whole request. Its advertised 1.05-million-token context window therefore should not be mistaken for a promise that very large inputs remain at the base price. Cache writes, longer retention, fast processing and external tools may add further costs.
For production agents, a smaller prompt with a targeted file lookup may be more economical than repeatedly sending a full document archive or an extensive tool history. A useful pre-launch check is to log input tokens, cached tokens, output tokens, tool charges, completion time and whether each job actually passed review.
Who can use GPT-6.1 Terra today?
OpenAI is making Terra available through ChatGPT Work and Codex to eligible paid users, including Plus, Pro, Business, Enterprise and Edu plans, subject to rollout and workspace controls. In ChatGPT Work, users can select it from the model choices in Advanced where available. It is not a separate model option in ordinary ChatGPT Chat.
Developers can call gpt-6.1-terra through the OpenAI API. Reasoning levels run from none to max, with medium as the default. The model supports text and image inputs, but does not natively accept audio or video. Tool calling requires the Responses API, even though the model can be used through other supported text endpoints.
Computer-use permission is a separate deployment decision from model selection. OpenAI says Terra meets its Critical cybersecurity capability threshold and carries over the safeguards developed for Astra. That does not make an agent with access to external systems immune to manipulation. An agent that can view a screen isn’t automatically safe to submit forms, change customer records, or execute commands. Limit tool access, require approval for consequential actions and treat instructions embedded in external documents and webpages as untrusted. For a practical example of constraining agent integrations, see DIY AI’s OpenCode MCP permissions and setup guide.
DIY AI’s assessment: a strong candidate for repetitive agent work
GPT-6.1 Terra is most interesting as a default for high-volume tasks where interpreting a screen, following instructions and working quickly matter more than maximum reasoning depth. Invoice processing, catalogue matching, document checks and structured coding changes are reasonable starting points for an evaluation, not guaranteed use cases.
Early developer reactions raise a fair question: why choose Terra rather than the cheaper GPT-6 Luna or the more capable GPT-6.1 Sol? The clearest reason would be a workload where Luna needs too many retries, but Sol’s extra reasoning seldom changes the outcome. That case needs measurement, not assumption.
Teams already using a stronger model should trial Terra against a representative batch of their own tasks. Record the share completed correctly on the first attempt, median completion time, retries, human intervention and total cost. Keep Sol or Astra available for genuinely difficult cases, rather than switching every workflow on the strength of a single benchmark.
The launch moves substantial computer-use capability into a cheaper model tier. Whether that produces meaningful savings will depend less on the $1 headline than on how much work Terra can finish accurately without supervision. OpenAI provides the full benchmark and launch details in its GPT-6.1 Terra announcement.
