OpenAI released GPT-6.1 Terra on October 26, and its API rates are half of what the previous Terra charged. Developers pay $1 per million input tokens and $5 per million output tokens, down from $2 and $12 for GPT-5.6 Terra, according to the model's developer page. The API is the pay-as-you-go route companies and programmers use to build the model into their own apps, separate from a ChatGPT subscription. Tokens are the chunks of text that usage is metered and billed in.
In ChatGPT, Terra is available in Work and Codex on the Plus, Pro, Business, Enterprise and Edu plans, but not in regular chat. In its announcement, OpenAI says Terra matches GPT-6 Astra, its most capable model, on a benchmark of professional tasks at a tenth of the price per token.
What it costs in practice
Input drops from $2 to $1 per million tokens, a 50% cut, and output from $12 to $5, a 58% cut. Text that repeats with every request, such as a fixed set of instructions, can be cached, meaning it is kept in temporary memory instead of being processed from scratch, and that cached input falls from $0.20 to $0.05 per million.
Terra sits between Sol and Luna in OpenAI's lineup. On OpenAI's pricing page, it costs half as much as GPT-6.1 Sol ($2 and $10) and a tenth as much as GPT-6 Astra ($10 and $50). GPT-6 Luna is still cheaper at $0.10 and $0.50, but it is aimed at narrower, repetitive jobs.
Take a customer support assistant that handles 10,000 questions a month, reading 2,000 tokens of documentation and writing a 500-token reply each time. OpenAI's rule of thumb is that 100 tokens come to about 75 English words, so that is roughly 1,500 words in and 375 words out per question. On Terra, the month costs $45: $20 for 20 million input tokens and $25 for 5 million output tokens. The same workload ran $100 on GPT-5.6 Terra. It costs $90 on GPT-6.1 Sol and $450 on GPT-6 Astra.
Real bills tend to come in higher. Reasoning models think before they answer, and OpenAI's reasoning guide says those hidden reasoning tokens are billed as output, so asking for more effort makes every reply more expensive. Requests over 272,000 input tokens are billed at twice the input rate and 1.5 times the output rate. Text in other languages also tends to use more tokens per word than English.
OpenAI's benchmark claims
Apart from the Artificial Analysis figures at the end, every comparison in this section comes from OpenAI, and none has been independently reproduced yet. Most depend on reasoning effort, an API setting that controls how long the model thinks before replying. It runs from none to max, and higher settings usually buy better answers with more tokens and more waiting.
The headline test is Agents' Last Exam, a set of professional tasks the model completes by operating desktop apps. At medium effort, Terra scores 15 percentage points higher than GPT-5.6 Terra at about a third of its predecessor's cost per task. On the same benchmark it ties GPT-6 Astra, and OpenAI says it finishes tasks more than three times faster than both Astra and Sol.
For coding, Terra solves 50.3% of Terminal-Bench 4.0 tasks at max effort, more than double the 21.5% GPT-5.6 Terra managed. The benchmark covers jobs done in the terminal, the text window developers use to run commands. OpenAI still points people to Sol and GPT-6 Astra for the hardest coding work.
OpenAI also says Terra makes fewer mistakes. On an internal set of questions designed to trip models up, the error rate at low effort falls from 53.6% with GPT-5.6 Terra to 16.7%, and at max effort it lands at 4.3%. An outside measurement points the same way. Artificial Analysis, as reported by The Decoder, found that Terra makes up an answer on 38% of the questions it cannot answer, down from 88% for GPT-5.6 Terra.
The same firm gives Terra 48 points on its Intelligence Index, against 52 for GPT-6.1 Sol and 42 for GPT-5.6 Terra. At max effort, it puts the cost of each task in its test suite at $0.40 for Terra and $1.40 for GPT-5.6 Terra.
Where you can use it
ChatGPT users pick Terra from the model list under Advanced in the desktop app or in ChatGPT Work on the web, according to the ChatGPT changelog. It also works in Work on mobile and in Codex, OpenAI's coding agent. It is not in regular chat, and the changelog does not mention the Free or Go plans. Enterprise access depends on what each company's admins turn on.
In the API the model is called gpt-6.1-terra. It takes text and images, replies in text and accepts up to 1,050,000 tokens of context, meaning the material it can work with in a single request. By OpenAI's rule of thumb, that is about 790,000 English words.
Safety and open questions
OpenAI rates Terra at the Critical level for cybersecurity in its Preparedness Framework, the scale the company uses to rate the risks of its models. Critical is the point at which a model could find and exploit flaws in hardened systems with little human help. It ships with the same safeguards as GPT-6 Astra, the model that led OpenAI to pause its largest training run.
OpenAI's own evaluations also say Terra stays inside its assigned task more often than GPT-5.6 Terra and is more likely to tell users about restrictions it runs into. That is where GPT-6.1 Astra fell short before it was canceled.
If you already pay for the API, the useful test is to rerun one of your own jobs on Terra and compare the cost of the finished task. A cheaper model that needs several attempts can end up costing more than one that gets it right the first time.
The conversation starts here
Sign in with a supporter account to comment. Sign in
Nobody has commented yet. Want to go first?