In the past few hours OpenAI has released GPT-6.1 Earththe mid-range model that replaces the GPT-5.6 Terra with a halved price compared to its predecessor: 1 dollar per million tokens in input and 5 dollars per million in output. Terra is the second of the four classes in the OpenAI range, after Luna and before Sol and Astra, in order of increasing price and capacity. The template is available to users Plus, Pro, Enterprise and Edu in ChatGPT Work and in Codex, and for developers via API with the identifier gpt-6.1-earth.
In the announcement, OpenAI defines Terra as “a fast and efficient model that combines detailed visual reasoning and precise computer use with clearer and more natural writing”, designed for the communication and computer work tasks that fill the day: from drafting proposals to interpreting technical drawings, from filling out forms to video editing. For computer use we mean the ability to look at the screen, understand what it shows and act accordingly, clicking, typing and switching from one app to another. For example, a company can entrust Terra with a flow of incoming invoices: the model reads each document and transcribes the data into the accounting software.
Developer pricing and specs
Standard rates are $1 per million token input And $5 per million in outputhalf that of GPT-5.6 Terra. Input tokens read from the cache cost 5% of the regular rate, i.e $0.05 per million. Requests in batch or with Flex mode cost 50% less than the standard rate, while fast mode costs double. For prompts over 272,000 input tokens, the price list applies double the input and cache rates and 1.5 times the output rates, with the surcharge extended to the entire request, while regional processing, where available, adds 10%.
The parameter reasoning.effort accepts six values: none, low, medium (default), high, xhigh, and max. The model produces up to 128,000 tokens in output and has a knowledge cutoff set at June 1, 2026. It supports function calling and structured outputs, among the available tools there are image generation and code interpreter, and to use the tools you need the Responses API.
Performance and security according to OpenAI
On Agents’ Last Exama test of complex professional tasks that includes converting architectural drawings into 3D models and compiling, for a restoration team, the time code log of an animated film viewed frame by frame, Terra scores at the medium reasoning level 15 percentage points more of GPT-5.6 Terra, at about a third of the cost. According to OpenAI the model ties GPT-6 Astra on the same test with a tenth of the price per token. Terra also completes the homework over three times faster of GPT-6 Astra and GPT-6.1 Sol.
On OSWorld 2.1which measures the ability to use desktop applications in multi-step workflows, Terra surpasses the best score achieved by GPT-5.6 Terra at the medium level with a significantly lower cost per task. On MMMU-Provisual reasoning test, beats its predecessor at every level tested and comes close to the results of GPT-6 Astra and GPT-6.1 Sol. In comparison with Claude Opus 5.5 of Anthropic, OpenAI claims that Terra completes tasks at about one-ninth the cost at the maximum reasoning level; the results come from OpenAI testing and have not been independently verified.
In OpenAI’s internal test, based on prompts designed to mislead, the error rate at the low reasoning level fell from 53.6% of GPT-5.6 Terra to 16.7%almost 37 percentage points less. At the maximum level the drop is 18 points and brings the rate to 4.3%close to the values that OpenAI indicates for GPT-6 Astra and GPT-6.1 Sol. Compared to its predecessor, Terra also stays within the assigned task more often and better signals its limits to users.
On Terminal Bench 4.0which measures how well coding agents solve command-line problems, such as fixing a web page whose layout shifts during loading, Terra solves the 50.3% tasks at the highest reasoning level. GPT-5.6 Earth stops at 21.5%so the new model has more than double the success rate. For more complex programming tasks OpenAI continues to recommend GPT-6.1 Sol or GPT-6 Astra.
Earth reaches level “Critical” for cybersecurity capabilities according to the Preparedness Framework, the system with which OpenAI classifies the risky capabilities of its models. The model inherits the same protections as GPT-6 Astra. According to OpenAI, it also offers greater resistance to prompt injection.
