OpenAI has once again raised the bar for the cost-effectiveness of large models.
Just now, OpenAI officially released its next-generation model, GPT-6.1 Terra, which boasts faster response times, enhanced visual understanding and computer operation capabilities, as well as lower usage costs.
Image | Blog https://openai.com/index/introducing-gpt-6-1-terra/
Compared to the previous-generation GPT-5.6 Terra, the new model’s standard API pricing has been cut by 50%, while delivering significant improvements in complex task performance, factual accuracy, and programming capabilities. Notably, in certain specialized task benchmarks, GPT-6.1 Terra has already approached the performance of the flagship GPT-6 Astra, yet its token pricing is only one-tenth of the latter’s.
OpenAI positions it as a highly efficient model for high-frequency work scenarios, aiming to enable more enterprises and developers to integrate AI Agents into their daily workflows at a lower cost.
Performance nearly matches Sol, with task execution speed three times faster.
The most noteworthy change in GPT-6.1 Terra is the further integration of visual understanding and computer manipulation capabilities.
In the past, AI could understand images and analyze documents, but to enable it to truly operate computers and perform complex tasks across applications, substantial model capabilities and high inference costs were still required.
Today, GPT-6.1 Terra is already capable of handling a wide range of professional tasks, from reading technical blueprints and filling out forms to editing videos and operating desktop software.
For example, users can have it generate 3D models based on architectural blueprints, or review animated films frame by frame to create time-stamped shot logs for the repair team. Companies can also use it to extract information from incoming invoices and automatically enter the relevant data into their accounting software.
In the Agents’ Last Exam test, which measures AI agents’ ability to perform professional tasks, GPT-6.1 Terra delivered a remarkably strong performance.
At a medium reasoning intensity, the new model’s score improved by 15 percentage points compared to GPT-5.6 Terra, while the cost per task was reduced to about one-third.
More notably, in the same test, GPT-6.1 Terra has already reached the performance level of GPT-6 Astra, while the standard API’s token price is only one-tenth of the latter’s.
Compared to Anthropic’s Claude Opus 5.5, at maximum reasoning intensity, GPT-6.1 Terra achieves more than three times the token efficiency, with a token price only one-quarter of Claude’s, resulting in a final cost approximately nine times lower for comparable tasks.
In terms of speed, GPT-6.1 Terra completes tasks more than three times faster than both GPT-6 Astra and GPT-6.1 Sol.
Test results shared by developers show that GPT-6.1 Terra can generate up to 450 tokens per second on a single request, and under low inference intensity, it may even reach 600 tokens per second—though the latter has not yet been verified.
The intelligence index ranking by the third-party evaluation firm Artificial Analysis also provides an interesting benchmark: GPT-6.1 Terra (Max) scored 48 points, surpassing all Chinese open-weight models except Kimi K3.1 Pro, which scored 55 points.
Additionally, in the MMMU-Pro test, which assesses visual reasoning capabilities, GPT-6.1 Terra outperformed the previous-generation model across all levels of reasoning intensity, achieving results comparable to those of GPT-6 Astra and GPT-6.1 Sol—at a significantly lower cost.
In OSWorld 2.1, which is specifically designed to test AI’s capabilities in desktop applications, the new model, even when using only medium inference intensity, has already outperformed GPT‑5.6 Terra at its highest inference intensity, while significantly reducing the cost per task.
Actual experience also confirms some improvements: DesignArena enabled two generations of Terra to recreate Lisbon’s Belém Tower at sunset, with version 6.1 producing more detailed architecture and refined light-and-shadow effects on the water surface. Although the rendering time is nearly four times that of the previous generation and token consumption has increased by 48%, the per‑render cost has dropped from $0.077 to $0.047.
Meanwhile, atomic.chat used the same prompts to have two models develop games with eight agents each: Terra completed the task in just 6 minutes and at a cost of $1.06, while Claude Sonnet 5.5 took about an hour and $18.10 but produced a noticeably superior game.
Moreover, based on existing tests, GPT-6.1 Terra still falls short of Claude Opus 5.5 in terms of final output quality on highly challenging tasks. However, given its lower price and faster execution speed, it still has considerable room to thrive. After all, in real-world scenarios, many tasks only require AI to get the job done, making it unnecessary to always deploy the most expensive flagship model.
Hallucination rate drops significantly, and programming performance doubles.
In addition to visual understanding and computer operations, GPT-6.1 Terra has also shown equally significant improvements in factual accuracy and programming capabilities.
OpenAI stated that this year it has been focusing on addressing the issues of model hallucinations and factual inaccuracies.
In an internally designed, high-difficulty fact-accuracy test, GPT-5.6 Terra recorded an error rate as high as 53.6% under low reasoning intensity, whereas GPT-6.1 Terra has reduced this to 16.7%, a decrease of 36.9 percentage points.
When the inference intensity is increased to the highest setting, the new model’s error rate drops further to 4.3%, approaching the levels of GPT-6 Astra and GPT-6.1 Sol.
For AI agents that need to process contracts, technical documents, and internal corporate data, a reduction in factual error rates is especially critical. After all, a model’s ability to complete tasks is only the baseline; whether it can consistently maintain accuracy throughout execution directly determines whether users can confidently entrust more work to the AI.
In terms of programming capabilities, GPT-6.1 Terra has also seen a significant improvement.
In Terminal-Bench 4.0, which tests AI agents’ command-line programming capabilities, GPT-6.1 Terra achieved a task resolution rate of 50.3% at the highest reasoning intensity, whereas the previous-generation GPT-5.6 Terra scored only 21.5% in public benchmarks.
In other words, the new model’s programming test scores have already surpassed twice those of the previous generation.
In the scientific research programming benchmark Terminal-Bench-Science 0.1, GPT-6.1 Terra even outperformed the higher-ranked GPT-6.1 Sol.
The test includes scientific research tasks such as analyzing cell division in time-lapse videos using code. OpenAI stated that GPT-6.1 Terra’s score is less than 5 percentage points behind Claude Opus 5.5, yet its estimated cost is less than one-fifth of the latter’s.
Of course, OpenAI has also made it clear that for the most complex software engineering and deep research tasks, GPT-6.1 Sol and GPT-6 Astra remain the more suitable choices.
API prices slashed by half, officially entering ChatGPT Work and Codex
Compared to the performance boost, the price adjustment for GPT-6.1 Terra may be more appealing to developers.
According to information released by OpenAI, the standard API pricing for the new model is as follows.
The input price is 1 USD per million tokens.
The cached input price is $0.05 per million tokens.
The output price is $5 per million tokens.
Compared to GPT-5.6 Terra and GPT-6.1 Sol, the standard API pricing for GPT-6.1 Terra has been reduced by 50%. When compared with the flagship model GPT-6 Astra, both its input and output token prices are only one-tenth of the latter’s.
At the product level, GPT-6.1 Terra is now available to ChatGPT Plus, Pro, Enterprise, and Edu users, and can be used in ChatGPT Work and Codex.
OpenAI stated that, with the same usage quota, users can complete more tasks using GPT-6.1 Terra than with GPT-5.6 Terra, and the improvement in quota efficiency is even more pronounced compared to GPT-6.1 Sol and GPT-6 Astra.
Developers can also invoke this model via the OpenAI API; the model is named gpt-6.1-terra.
In terms of security, OpenAI stated that GPT-6.1 Terra has reached the Critical risk level for cybersecurity capabilities within its Preparedness Framework, and therefore continues to implement the relevant security safeguards used in models such as GPT-6 Astra.
Meanwhile, the new model inherits the alignment improvements of the GPT-6 series, showing enhancements over the previous generation in adhering to user‑authorized boundaries, proactively communicating its capability limits, and resisting malicious prompt injection.
Looking back at OpenAI’s product updates this year, from the flagship model Astra to Sol, which emphasizes general capabilities, and now to Terra, which focuses on efficient execution, the division of labor among these products is becoming increasingly clear.
Especially as AI agents accelerate their adoption in real-world work scenarios, model competition has shifted from merely pursuing higher benchmark scores to evaluating how much it costs and how long it takes to complete a given task, as well as whether the model can perform consistently and reliably.
In the past, when companies sought to use AI to automate large volumes of office tasks, they often had to repeatedly weigh model capabilities, speed, and cost.
The emergence of GPT-6.1 Terra is reducing the cost of this trade-off.
When a model that can read screens, operate software, and comprehend complex documents has already achieved near-flagship-level capabilities yet requires only one-tenth of the token cost, AI Agents gain the practical conditions to expand from handling a small number of high-value tasks to widespread everyday applications.
For OpenAI, making increasingly powerful AIs more affordable may be just as important as releasing its next-generation flagship model.
#Welcome to follow the iFanr official WeChat account: iFanr (WeChat ID: ifanr), where we bring you the latest and most exciting content第一时间.