OpenAI Releases GPT-6.1 Terra: A Faster, Cheaper, and Stronger Model for Agent Tasks

Launch and Key Characteristics

OpenAI announced GPT-6.1 Terra on October 26, 2026. The new model is described as a fast, efficient medium-sized model. It combines detailed visual reasoning, precise computer use, and clearer, more natural writing. It is the combination of these three directions in a single model that defines the core identity of GPT-6.1 Terra. The model is designed not simply to answer faster, but to improve seeing, acting, and writing at the same time.

a fast, efficient medium-sized model combining detailed visual reasoning and precise computer use with clearer, more natural writing.

GPT-6.1 Terra succeeds GPT-5.6 Terra. OpenAI presents this transition as a significant step forward from the previous generation. The update matters especially for the Terra family, because last month, when GPT-6 models were released for Sol and Luna, Terra was not included in that lineup. GPT-6.1 Terra now closes that gap and brings the Terra line into the new generation. Continuity within the model family is thus restored, and users gain updated capabilities under a familiar name.

The announcement of the new model is aimed at several audiences at once. ChatGPT Work and Codex users, as well as developers working through the API, encounter the same model name on the same day. This approach presents the model not merely as a research announcement, but as a practical working tool. The combination of speed, visual reasoning, and computer use is especially visible in multi-step tasks, where the model must first see, then decide, and then express the result in precise text. More detail is given in the OpenAI blog.

Pricing: Half the Price of the Previous Terra

The standard API prices for GPT-6.1 Terra are set 50 percent lower than the prices of GPT-5.6 Terra and GPT-6.1 Sol. One million input tokens cost $1, one million cached input tokens cost $0.05, and one million output tokens cost $5. These figures mean the model is twice as cheap as the previous Terra generation in price terms. At the same time, the new model also costs half as much as GPT-6.1 Sol. The price difference is especially noticeable in projects with a large volume of requests.

The price drop comes together with the model itself becoming faster and more efficient. Speed and cheapness are usually seen as a conflicting trade-off, but this time OpenAI reports progress in both directions. The low input price reduces cost pressure when working with long documents, large codebases, and broad contexts. The very low cached input price offers additional savings when working with repeated contexts. Because the output price is also below the previous level, long answers and multi-step agent tasks cost less.

The pricing policy becomes one of the main factors in choosing GPT-6.1 Terra. Within the same budget, a user can complete more tasks or complete the same volume of work at a lower cost. This matters especially for agent systems, because agents call the model several times within a single task. When the price of each call falls, the total project cost falls too. From this perspective, GPT-6.1 Terra takes the price-capability balance to a new level compared with the previous Terra generation.

Significant Gains on Agents Last Exam

On the Agents Last Exam benchmark, at medium reasoning effort, GPT-6.1 Terra scored 15 percentage points higher than GPT-5.6 Terra. This gap was achieved at roughly one-third of the cost. In other words, the new model not only earned a higher score, but reached that result at a much lower cost. The simultaneous improvement in score and cost shows that the model's efficiency rose on both measures. Compared with the previous generation, this gap is especially clear in large-scale agent tasks.

On the same benchmark, GPT-6.1 Terra matched the score of GPT-6 Astra, but its price per token was one-tenth of Astra's price. This comparison shows the new model standing on a higher rung in price-performance terms. An equal score at a tenfold lower price makes the choice for large-scale use much clearer. The chance to get Astra-level results at a Terra price becomes one of the factors defining the model's market position. According to OpenAI, this balance reflects the very essence of the medium-sized model concept.

The speed gap was also large. GPT-6.1 Terra completed tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol. Speed is decisive in agent systems, because each task consists of several steps and the delay at each step adds to the total time. Three times the speed means more tasks in the same time, or a single task in a significantly shorter time. Speed therefore defines the model's practical advantage, alongside price and accuracy. Full results are described in the OpenAI blog.

Errors Down Sharply

In OpenAI's internal benchmark, GPT-6.1 Terra's error rate on error-inducing prompts fell sharply. At low reasoning effort, the error rate fell from 53.6 percent for GPT-5.6 Terra to 16.7 percent for GPT-6.1 Terra. That is almost a threefold decline. It matters especially that the model's tendency to err fell markedly even in the low-effort mode, since the low-effort mode is usually chosen for speed and cheapness. Now, even in this mode, reliability is much higher than in the previous generation.

At maximum reasoning effort, the error rate fell by 18 percentage points, to 4.3 percent. This figure shows that the model became much more resistant to error-inducing prompts even in its most thorough mode. The gap between low and maximum effort remaining is expected, but the decline in both modes indicates that the model's overall stability has risen. Because error-inducing prompts are specifically designed to mislead the model, the improvement in this benchmark directly affects practical reliability.

The lower error rate matters both in agent tasks and in ordinary text conversation. With fewer errors, less time goes to checking and correcting results. At the same time, the chance that the model goes off in the wrong direction falls, which prevents errors from spreading in chains in multi-step tasks. The gain in this metric is therefore not only a numerical improvement, but a quality change felt in everyday use.

More Than Double the Score on Terminal-Bench 4.0

On the Terminal-Bench 4.0 benchmark, at maximum reasoning effort, GPT-6.1 Terra solved 50.3 percent of tasks. That is more than double the announced 21.5 percent result of GPT-5.6 Terra. Terminal tasks require executing commands in a computer environment, working with files, and solving problems step by step. The result in this benchmark therefore directly reflects the model's computer-use ability. A 50.3 percent result means every second task was completed successfully.

Compared with the previous generation, the gap is especially stark. The move from 21.5 percent to 50.3 percent shows that the model's capabilities in the terminal environment have reached a qualitatively new level. This growth is not a small improvement, but a more than doubling of the probability of solving tasks. In practical terms, this means developers and agent systems can rely on the model more for terminal-related tasks. Unresolved tasks still remain, of course, but the success share has risen markedly.

The Terminal-Bench result also aligns with the description of the model as capable of precise computer use. The model must see the task, choose the necessary command, and, having evaluated the result, move to the next step. This process requires a chain of multi-step actions rather than a one-off answer. GPT-6.1 Terra's result in this benchmark confirms that the ability to manage exactly such a chain has grown. The Terminal-Bench 4.0 result therefore became one of the main pieces of evidence in the model's presentation.

Safety Level and Availability

GPT-6.1 Terra reached the Critical level for cybersecurity under OpenAI's Preparedness Framework. This level means the model's cybersecurity capabilities have passed a dedicated assessment and fall into a category requiring heightened attention. OpenAI disclosed this information openly together with the model announcement. Open disclosure of the safety level gives a clear picture both of the model's capabilities and of the precautions applied to it. This aspect was therefore noted as an important part of the presentation.

From October 26, 2026, the model became available to all Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex. For developers, the model was offered through the API under the name gpt-6.1-terra. Launching on the same day in both ChatGPT Work and Codex and through the API ensured the model reached a broad audience at once. Users on different plans see the same model name and use the same core capabilities. This speeds up both the spread of information about the model and practical testing.

The breadth of availability, combined with the model's speed and price advantages, widens the application range of GPT-6.1 Terra. Small teams and individual developers as well as large organizations can use the same model. The API name being precisely defined as gpt-6.1-terra simplifies the integration process. The announcement was thus presented not only as research news, but as a product that can be used immediately.

Terra's Place in the GPT-6 Family

The release of GPT-6.1 Terra clearly defined the place of the Terra line in the GPT-6 generation. Last month, when GPT-6 models were released for Sol and Luna, Terra was left aside. Now Terra too has joined the new generation, under the 6.1 designation. This sequence shows that there can be a time gap between model families, but each line is updated in its turn. For Terra, this update is presented as a major step after the previous generation.

Taken together, price, speed, and benchmark results show GPT-6.1 Terra as a medium-sized model aiming to combine high results with low cost. It demonstrated much higher scores than GPT-5.6 Terra, offered a result equal to GPT-6 Astra at a lower price, and turned out both cheaper and faster than GPT-6.1 Sol. These three comparisons clearly draw the model's place in the family. Each comparison shows an advantage on a separate measure, and together they form an overall balance.

Thus, the presentation of GPT-6.1 Terra brings together changes in several directions: the replacement of the previous Terra generation, the halving of price, high results in agent benchmarks, the sharp fall in the error rate, and a more than doubled result on terminal tasks. All of these changes were announced on October 26, 2026, and from that day the model became available through ChatGPT Work, Codex, and the API. The model's further assessment will now depend on whether these announced figures are confirmed in practical use.