OpenAI cuts the price of Terra in half and brings its agents closer to the most powerful models

GPT-6.1 Terra improves software use, reduces factual errors, and costs half as much as its predecessor via the API.

A model capable of running professional software, understanding complex visual documents, and completing tasks more than three times faster than some models that are nevertheless more powerful. With GPT-6.1 Terra, OpenAI is prioritizing computer agent efficiency while cutting the pricing of its mid-tier model by half.

The evolution goes beyond simple text generation acceleration. Terra is designed to understand what appears on a screen, act within applications, interpret technical documents, and carry out multi-step operations. It is now available in ChatGPT Work, Codex, and the OpenAI API, with standard pricing of $1 per million input tokens and $5 per output token. Agents Capable of Working Across Multiple Software Applications

Reading an invoice and entering its data into accounting software, filling out a form, converting an architectural drawing into a 3D model, or analyzing an animated film frame by frame: the examples highlighted by OpenAI involve tasks requiring much more than a text response.

GPT-6.1 Terra combines visual understanding, reasoning, and computer interface usage. When integrated into an agentic environment with the necessary tools and permissions, the model can interpret an interface, determine the operations to perform, and interact with various applications.

In ChatGPT Work, these capabilities fit into workflows that make it possible to entrust the AI with extended tasks involving documents, software, and connected services. In Codex, they can be used for development and testing operations that require computer usage.

The difference also lies in the level of the model used. Complex computer agents have long benefited from the most expensive models, particularly when they had to chain together numerous visual observations and actions. Terra seeks to perform a portion of these operations at a lower cost and faster processing time. Fifteen Points Better on Professional Tasks

On Agents' Last Exam v1, a benchmark measuring the ability of agents to perform professional tasks in software, GPT-6.1 Terra improves upon the result of GPT-5.6 Terra by 15 percentage points at medium reasoning effort.

Tasks notably include building three-dimensional models from architectural documents or producing timestamped logs from video sequences.

OpenAI indicates that this progress comes with a cost per task approximately three times lower than that of GPT-5.6 Terra.

At this reasoning level, GPT-6.1 Terra also achieves performance comparable to that of GPT-6 Astra, while featuring a price per token ten times lower. Furthermore, the model completes tasks more than three times faster on average than GPT-6 Astra and GPT-6.1 Sol, according to evaluations published by OpenAI.

These comparisons distinguish between two parameters that are often confused: the price of tokens and the cost of a task actually completed. A model with cheaper tokens may consume more tokens or make multiple attempts. The economic outcome therefore also depends on its efficiency during execution.

At maximum reasoning effort, OpenAI estimates that Terra uses more than three times fewer tokens than Claude Opus 5.5 for the compared tasks. Taking the pricing into account, the company claims a cost per task approximately nine times lower. These results come from OpenAI's measurements and do not demonstrate general superiority across all use cases. Factual Errors Drop from 53.6% to 16.7%

The improvement also applies to response reliability.

On an internal test set composed of questions prone to causing errors, GPT-5.6 Terra produced at least one factual error in 53.6% of responses at low reasoning effort. With GPT-6.1 Terra, this rate drops to 16.7%, representing a 36.9 percentage point decrease.

At maximum effort, the error rate of the new version reaches 4.3%, compared to 22.3% for the previous generation. The results then approach those of the GPT-6 Astra and GPT-6.1 Sol models in this evaluation.

These figures do not correspond to a universal hallucination rate. They originate from an internal benchmark on difficult prompts, whose results cannot be directly extrapolated to all conversations or professional documents.

On MMMU-Pro, a benchmark dedicated to visual reasoning, OpenAI also reports progress across all reasoning levels compared to GPT-5.6 Terra. The highest results approach those of Astra and Sol, at a lower cost.

This understanding of visual content is particularly relevant when an agent needs to read a table, distinguish elements of a technical diagram, or understand an interface before performing an action. Programming Capabilities See Strong Progress

Terra also improves its results on software development tasks.

On Terminal-Bench 4.0, which evaluates problem-solving in a command-line environment, GPT-6.1 Terra successfully completes 50.3% of tasks at maximum reasoning effort. GPT-5.6 Terra achieved 21.5% in previously published results.

This represents a progress of 28.8 points and a score more than twice as high.

The benchmark includes concrete development operations, such as fixing a web interface whose elements shift position while loading.

Results also extend to program-assisted scientific research. On Terminal-Bench-Science 0.1, Terra surpasses GPT-6.1 Sol at a lower cost and finishes within five points of Claude Opus 5.5, for less than a fifth of its estimated cost.

However, OpenAI continues to recommend GPT-6.1 Sol and GPT-6 Astra for the most challenging programming tasks and research requiring deep reasoning. A Cheaper Model to Scale Operations

The new pricing structure represents one of the most direct changes for developers.

The GPT-6.1 Terra API is priced at $1 per million input tokens, $0.05 for cached inputs, and $5 per million output tokens. These rates represent a 50% reduction compared to the standard prices announced for GPT-5.6 Terra and GPT-6.1 Sol.

This difference can have a significant impact on applications making numerous calls to the model. An agent tasked with verifying invoices, updating a database, or