OpenAI Launches GPT-6.1 Terra: API Price Halved, Optimized for Agent Tasks

OpenAI has officially released GPT-6.1 Terra, a fast model built for computer-use and vision capabilities. Priced at $1 per million input tokens and $5 per million output tokens, OpenAI says it delivers a major boost in cost-effectiveness for Agent tasks—completing the same task for just $0.29, compared with GPT-6 Astra's $4.82 for a medium-intensity task.

Core Positioning: A Speed-Focused Model Built for Agent Tasks

GPT-6.1 Terra's positioning is clear: it isn't chasing all-around flagship performance, but focuses on the high-frequency, fast-response scenarios within agent workflows. According to MarkTechPost, Terra matches GPT-6 Astra on the "Agents' Last Exam" benchmark under medium-intensity settings, yet the per-task cost differs by orders of magnitude.

For agent systems that need to call the model in loops and execute multi-step operations (such as browsing web pages, filling out forms, and clicking interface elements), the cost of each call is significantly amplified as the number of steps grows. Terra pushes the per-million-token price down to the $1/$5 range and measures performance by "task cost" rather than simple per-token pricing—a pricing approach that itself reflects an understanding of long-chain agent scenarios.

Agents' Last Exam is a benchmark specifically designed to evaluate AI agents' end-to-end performance in real, multi-step workflows, distinct from traditional static Q&A or single-turn reasoning evaluations. It typically requires models to autonomously plan, invoke tools, operate interfaces, and complete open-ended tasks in simulated or real environments, with scoring dimensions including task success rate, number of steps to completion, and resource consumption. The so-called "medium-intensity setting" generally refers to task complexity, allowed interaction rounds, or time budgets at the middle tier of the benchmark, enabling controlled comparisons across models of different capability tiers. Therefore, "matching scores" on this benchmark means the two models are indistinguishable in task completion quality, and it is the difference in "per-task cost" that truly reveals efficiency and the economics of scaled deployment.

Pricing Structure: Dual Advantages in API Unit Price and Task Cost

In terms of API pricing, GPT-6.1 Terra costs $1 per million input tokens and $5 per million output tokens. The headline mention of "Halving GPT-6.1 Sol's API Price" indicates that Terra's pricing is roughly half that of the same-generation Sol model.

More notable is the task-level cost comparison: for the same medium-intensity Agents' Last Exam task, Terra's per-task cost is $0.29, versus $4.82 for GPT-6 Astra. This gap is not caused solely by the per-token price; it may also relate to the number of interaction rounds and output length the model requires to complete the task. In other words, Terra's efficiency advantage may be further amplified in multi-step agent workflows.

Vision and Computer-Use Capabilities Are the Focus

Terra is explicitly defined as a "fast computer-use and vision model." Computer-use means the model can directly understand screen captures and operate graphical interfaces—a key direction for current agent products. The addition of vision capabilities lets the model handle tasks such as interface element recognition and chart understanding, while "fast" points to latency-sensitive scenarios.

In practice, this means developers can use Terra to build automation tools that require real-time feedback—for example, automatically filling in backend systems, moving data across applications, and page inspection. Compared with pure text models, models with visual operation capabilities reduce their reliance on custom interfaces for such tasks.

Computer-use refers to a model understanding graphical user interfaces through visual input (such as screen captures) and outputting operation instructions like clicks, typing, scrolling, and dragging, thereby using software directly like a human rather than relying on pre-wrapped APIs. Such models typically combine a vision encoder with an action prediction head, advancing tasks step by step in an "observe—decide—execute" loop, and are sensitive to both latency and action precision. Their advantage lies in covering legacy systems without open interfaces and cross-application workflows, but they also face robustness challenges such as interface element recognition errors, dynamic layout changes, and error accumulation over long chains. For this reason, the stability of computer-use models in real production environments often requires additional validation.

What It Means for Developers: Agent Cost Structures Are Changing

Over the past year, one of the main bottlenecks for agent-type applications has been inference cost. A complex task may require dozens of model calls, and the high unit price of flagship models has made it difficult for many teams to deploy at scale. Terra's arrival pushes the cost of medium-intensity tasks below $0.3, meaning the same budget can support more concurrent tasks or longer automation chains.

One caveat: currently available public information covers only benchmark results and pricing. There is no further data on Terra's stability in real production environments, its robustness to complex interfaces, or how it compares head-to-head with computer-use models from other vendors. Developers should run small-scale validation against their actual workflows before getting started.

Market Signal: Tiered Model Strategies Are Becoming Mainstream

Terra's release continues OpenAI's tiered product approach: flagship models handle the highest-difficulty reasoning, while lightweight models handle high-frequency, scaled calls. The simultaneous mention of GPT-6.1 Sol and GPT-6 Astra shows that its product line has already formed multiple price tiers within the same generation. For developers, model selection is no longer just about "picking the strongest," but about weighing capability, latency, and cost.

From the perspective of the agent ecosystem, halved prices and a sharp drop in task costs may push more teams from "demo-level agents" toward "agents that can be deployed at scale." The real test is whether Terra can maintain a sufficiently high task success rate while keeping latency low.

One-line Summary

OpenAI releases GPT-6.1 Terra, focused on low-cost, high-speed computer-use and vision capabilities, targeting scaled Agent deployment.

Paragraph Summary

OpenAI has released GPT-6.1 Terra, a fast model built for computer-use and vision capabilities. Its API pricing is $1 per million input tokens and $5 per million output tokens, roughly half that of the same-generation GPT-6.1 Sol. OpenAI emphasizes task-level cost-effectiveness: on medium-intensity Agents' Last Exam tasks, Terra matches GPT-6 Astra's scores, with a per-task cost of just $0.29 versus Astra's $4.82. The article argues that this pricing approach—measured by "task cost" rather than simple per-token price—reflects an understanding of multi-step, long-chain agent scenarios and could lower the barrier to scaled deployment of agent applications, though stability and robustness in real production environments remain to be verified.