GPT-6.1 Terra cuts standard API token prices by 50% versus GPT-6.1 Sol, charging $1 per million input tokens and $5 per million output tokens in 2026. Released by OpenAI on October 26, 2026, the model targets office automation, scoped code repairs and high-volume agent workloads where cost per completed task matters more than winning every benchmark.
What is GPT-6.1 Terra?
GPT-6.1 Terra is an OpenAI reasoning model designed for screen-based workflows, concise writing from supplied material and bounded software fixes. OpenAI released the model as gpt-6.1-terra through the Responses and Chat Completions APIs on October 26, 2026, alongside access in supported ChatGPT Work and Codex plans.
The model has a 1.05-million-token context window, a maximum output of 128,000 tokens and a June 1, 2026 knowledge cutoff, according to OpenAI’s October 2026 model documentation. Those limits accommodate large repositories and document collections, but a large context window doesn’t guarantee that filling it is economical or accurate.
Through the Responses API, Terra supports computer use, web search, file search, hosted shell, code interpreter, structured outputs and function calling. That combination makes the model relevant to browser agents and desktop automation, an area also moving toward local AI agents on Windows 11.
OpenAI launched GPT-6.1 Terra for supported Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex on October 26, 2026. API availability matters more for the cost comparison, because developers can measure token consumption and completed-task rates directly.
How much does GPT-6.1 Terra cost in 2026?
GPT-6.1 Terra costs $1 per million standard input tokens, $0.05 per million cached-input tokens and $5 per million output tokens as of October 26, 2026. OpenAI prices Batch and Flex processing at 50% below those Standard rates, making asynchronous input and output cost $0.50 and $2.50 per million tokens.
The headline rate is easy to understand. Your invoice, however, depends heavily on output length because one million generated tokens cost five times as much as one million uncached input tokens under the 2026 Standard tariff.
| Model or processing tier | Input | Cached input | Output |
|---|---|---|---|
| GPT-6.1 Terra Standard | $1.00 | $0.05 | $5.00 |
| GPT-6.1 Terra Batch or Flex | $0.50 | $0.025 | $2.50 |
| GPT-6.1 Sol Standard | $2.00 | $0.10 | $10.00 |
| Claude Sonnet 5.5 Standard | $2.00 | $0.10 cache read | $10.00 |
Consider a 2026 workload processing 100 million uncached input tokens and producing 20 million output tokens. Terra Standard costs $100 for input plus $100 for output, or $200 total. The same token volumes cost $400 with GPT-6.1 Sol or Claude Sonnet 5.5 at their published Standard rates.
Batch or Flex would reduce the Terra calculation to $100 in 2026. OpenAI introduced optional 12-hour prompt caching on October 20, 2026, but writes using that retention period cost 2.5 times the standard input-token rate. Longer caching can still pay off when the same costly prefix is read repeatedly; infrequent reuse can erase the benefit.
Is GPT-6.1 Terra really half the cost of Sol and Claude?
GPT-6.1 Terra has token rates exactly 50% below GPT-6.1 Sol and Claude Sonnet 5.5 under their published Standard pricing on October 26, 2026. That price relationship does not prove a 50% reduction per completed workload because reasoning effort, retries, tool calls and generated-token totals can differ by model.
Cost per successful task is the better metric. Divide total model and tool spending by the number of jobs that satisfy your acceptance test, then include retries rather than quietly excluding them. A cheaper attempt becomes expensive if weak first-pass results trigger repeated calls or human repair.
OpenAI reported on October 26, 2026, that Terra completed Agents’ Last Exam tasks at medium reasoning effort more than three times faster than Astra and Sol. The company also said Terra scored 15 percentage points above GPT-5.6 Terra while costing about one-third as much per task. These are vendor-reported evaluation results, not independent production measurements.
At maximum effort, OpenAI reported in 2026 that Terra used tokens more than three times as efficiently as Claude Opus 5.5 and cost approximately nine times less per Agents’ Last Exam task. OpenAI did not publish an equivalent task-cost comparison against Claude Sonnet 5.5, so extending the ninefold claim to Sonnet would be unsupported.
My view is simple: Terra’s pricing is unusually attractive when your jobs have clear pass-or-fail checks. For open-ended professional work where quality is difficult to score, a small pilot tells you more than token prices ever will.
Where does GPT-6.1 Terra make the most sense?
GPT-6.1 Terra best fits repeatable office automation, high-volume API processing, browser control and scoped coding tasks in 2026. OpenAI recommends the model for screen-based workflows, writing grounded in supplied material and contained code fixes, while directing more complex professional and software projects toward GPT-6.1 Sol.
Good deployment candidates share four practical traits. The following checks identify workloads where Terra’s lower rates can translate into lower operating costs rather than merely cheaper individual calls:
- The task has a machine-checkable result, such as a passing test, valid schema or confirmed form submission.
- Prompts reuse a stable document, policy or code prefix often enough to benefit from cached-input pricing.
- The workflow can tolerate Batch or Flex scheduling and capture the 50% processing discount available in 2026.
- A human or automated rule can stop repeated tool calls when the agent enters a loop.
Examples include extracting structured fields from files, drafting short summaries from a supplied packet, repairing a failing unit test and completing a defined browser procedure. Research assistants such as Unriddle AI for document research and data tools such as Clara by Pythagoras AI illustrate the broader class of grounded workflows where input quality and verification matter as much as model choice.
Coding agents deserve closer scrutiny. OpenAI reported a 50.3% Terminal-Bench 4.0 score for Terra at maximum reasoning effort in 2026, compared with its published 21.5% result for GPT-5.6 Terra. Anthropic reported 70.6% for Claude Sonnet 5.5 in September 2026, but differences in evaluation settings prevent a clean model-to-model ranking.
Honestly, GPT-6.1 Terra only makes sense for advanced coding if your own repository tests confirm that the cheaper model finishes enough jobs. OpenAI still recommends Sol or Astra for its most complex coding tasks, despite Terra’s substantial generational gain.
What are GPT-6.1 Terra’s accuracy and security limits?
GPT-6.1 Terra still produces factual errors and requires task-level verification. OpenAI reported a 16.7% difficult-prompt factual-error rate at low reasoning effort and 4.3% at maximum effort on October 26, 2026, compared with 53.6% and 22.3% respectively for GPT-5.6 Terra.
The accuracy improvement is large in OpenAI’s evaluation, but maximum reasoning may consume more tokens and time than low effort. You should measure the cost of accepted answers at each effort setting rather than assume that maximum effort is always economical.
Computer control adds another failure mode: an agent can misunderstand a page, click the wrong control or repeat an action. Use confirmation gates before purchases, messages, account changes and destructive commands. Log screenshots, tool calls and final states so failed automation can be audited.
Security deserves equal attention. OpenAI’s October 26, 2026 safety addendum classified Terra as Critical for cybersecurity capability and High for biological and chemical capability, with the same safeguard stack used for Sol and Astra. Organizations exposing AI infrastructure should also account for threats such as malware targeting AI servers, which model safeguards alone don’t address.
The overlooked pitfall is authorization sprawl. A low-cost agent can attempt far more actions within the same budget, increasing exposure if browser sessions, shell access or API credentials have broad privileges. Give every workflow the narrowest permissions possible and set monetary, token and tool-call limits.
GPT-6.1 Terra FAQ
What is the GPT-6.1 Terra API model name?
The GPT-6.1 Terra API model identifier is gpt-6.1-terra, according to OpenAI’s October 26, 2026 release documentation. OpenAI provides access through the Responses and Chat Completions APIs.
What is GPT-6.1 Terra’s context window?
GPT-6.1 Terra has a 1.05-million-token context window and supports a maximum output of 128,000 tokens, according to OpenAI’s October 2026 model documentation.
What is GPT-6.1 Terra’s knowledge cutoff?
GPT-6.1 Terra has a June 1, 2026 knowledge cutoff. Web search can retrieve later information when the Responses API workflow enables that tool, but retrieved claims still require source checking.
Does GPT-6.1 Terra support computer use?
GPT-6.1 Terra supports computer use through the Responses API as of October 26, 2026. OpenAI also lists web search, file search, hosted shell, code interpreter, structured outputs and function calling among the supported capabilities.