OpenAI released GPT-6.1 Terra on Monday as a lower-cost model for visual reasoning, computer control and routine knowledge work. It is available now in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and through the API under the model ID gpt-6.1-terra.
The headline price is $1 per million input tokens and $5 per million output tokens, half the standard rates OpenAI charged for GPT-5.6 Terra and GPT-6.1 Sol. That makes the new model an appealing candidate for screen-based automations that process a large queue of similar tasks: reading invoices, matching product photos to catalog records, filling forms, reviewing technical drawings or moving information between applications.
But the cheapest-looking model is not automatically the cheapest deployment. Terra supports a 1.05 million-token context window, and OpenAI applies higher rates once an individual prompt exceeds 272,000 input tokens. Teams should compare completed-task cost at the reasoning setting and context size they actually plan to use.
GPT-6.1 Terra pricing and limits
The model documentation lists a 1,050,000-token context window, a maximum output of 128,000 tokens and a June 1, 2026 knowledge cutoff. It accepts text and images as input and produces text; native audio and video input are not supported.
| Item | Standard rate or limit |
|---|---|
| Input | $1 per million tokens |
| Cached input | $0.05 per million tokens |
| 30-minute cache write | $1.25 per million tokens |
| 12-hour cache write | $2.50 per million tokens |
| Output | $5 per million tokens |
| Context window | 1.05 million tokens |
| Maximum output | 128,000 tokens |
Prompts above 272,000 input tokens are charged at twice the input and cache rates and 1.5 times the output rate for the entire request. Batch and Flex processing cost 50% less than the standard tier, while Fast mode costs twice as much. Regional processing adds a 10% premium where offered.
Those modifiers can change architecture decisions. A workflow with a stable policy manual, tool schema and reference set can benefit heavily from the $0.05 cached-input rate after paying to write the cache. A workflow that repeatedly sends a different 300,000-token document crosses the long-context threshold and loses much of Terra’s apparent price advantage.
Where Terra fits in the GPT-6 lineup
OpenAI positions Terra below GPT-6.1 Sol and GPT-6 Astra for the hardest reasoning and coding jobs, but above the smaller GPT-6 Luna when visual interpretation and reliable interaction with software matter. The practical dividing line is the task, not simply the benchmark rank.
Terra is the model to test first when an agent must inspect a screen, understand an image or document, and then take a bounded action. Sol or Astra remains the safer starting point for ambiguous research, difficult software engineering and work where a failed reasoning step is more expensive than additional inference cost. Luna is the lower-cost option for narrow classification, triage and other high-volume tasks that do not need Terra’s stronger computer-use performance.
On OpenAI’s Agents’ Last Exam, which includes professional software tasks such as turning architectural drawings into a 3D model, Terra at medium reasoning scored 15 percentage points above GPT-5.6 Terra at roughly one-third of the cost. OpenAI also reports that it completed those tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol. These are provider-run evaluations, so production teams should treat them as a shortlist signal and replay their own representative tasks before migrating.
The improvement is not confined to screen control. At maximum reasoning effort, Terra solved 50.3% of Terminal-Bench 4.0 tasks, compared with GPT-5.6 Terra’s published 21.5%. OpenAI still recommends Sol or Astra for the most complex coding work.
Tool use requires the Responses API
Terra appears across several API endpoints, including Responses and Chat Completions, but its tool calling requires the Responses API. Supported tools there include web search, file search, image generation, code interpreter, hosted shell, computer use, MCP and skills.
That detail matters for migrations. An application can send ordinary text and image requests through a familiar endpoint, but an agent that clicks through software or calls business tools needs the Responses API path. Teams moving from another model should verify tool schemas, approval gates, retry behavior and how the agent recovers when an interface changes.
The model supports none, low, medium, high, xhigh and max reasoning effort, with medium as the default. Testing only the default can hide the real cost-performance curve. A useful evaluation matrix runs the same task set at low, medium and high effort, recording completion rate, human corrections, tokens, latency and tool calls. Max effort belongs in that test only when the extra reasoning has a plausible economic payoff.
What teams should test before switching
Start with workflows that are repetitive, visually grounded and easy to verify. Invoice entry is a better first deployment than an open-ended finance agent because the source fields and destination records can be checked. Product matching, form completion, document routing and screen-based quality assurance offer similarly clear pass-fail criteria.
- Measure cost per accepted task. Token price alone misses retries, long-context multipliers, reasoning tokens and human review.
- Separate perception from action. Log what Terra saw, the action it proposed and the action actually executed so failures can be traced.
- Keep destructive steps gated. Require approval before sending messages, submitting payments, deleting records or changing permissions.
- Test prompt injection. Put malicious instructions inside documents, webpages and images to check whether the agent abandons its assigned task.
- Re-run UI tests after application changes. Computer-use reliability can fall when buttons move, dialogs change or accessibility labels disappear.
OpenAI reports a large factuality improvement over GPT-5.6 Terra: on an internal set of difficult, error-inducing prompts, the error rate at low reasoning fell from 53.6% to 16.7%, while max effort reached 4.3%. That is meaningful progress, but it does not remove the need for validation in workflows that alter external systems.
The cybersecurity classification deserves attention
GPT-6.1 Terra meets OpenAI’s Critical cybersecurity threshold under its Preparedness Framework and inherits protections developed for GPT-6 Astra. OpenAI also reports stronger resistance to prompt injection and better adherence to task boundaries than GPT-5.6 Terra.
For buyers, the classification is not a reason to avoid ordinary use. It is a reason to treat broad tool access as privileged access. A fast, inexpensive model capable of operating software can scale mistakes and abuse as efficiently as legitimate work. Least-privilege credentials, isolated execution, detailed audit logs and human approval for consequential actions should be part of the deployment design, not features added after a pilot succeeds.
GPT-6.1 Terra’s most important change is economic: capable visual agents and computer-use workflows can now be tested at a much lower token price. The teams that benefit will be the ones that evaluate the whole system, including context length, caching, reasoning effort, tool fees, retries and oversight, rather than choosing a model from the price column alone.
Sources: OpenAI’s GPT-6.1 Terra announcement, model documentation and API pricing page.
