GPT-6.1 Terra arrives in Codex at a tenth of GPT-6 Astra's price

OpenAI added GPT-6.1 Terra to ChatGPT Work and Codex on October 26, for Plus, Pro, Business, Enterprise and Edu plans. OpenAI pitches it as a fast, efficient model for reading images and screens, precise computer use and plain writing. It is picked from the model list under Advanced, it is not offered in Chat, and Enterprise workspaces get it only where an administrator allows it.

The price is the news. The API model card lists Terra at $1 per million input tokens and $5 per million output, with cached input at $0.05. OpenAI's pricing page puts GPT-6.1 Sol at twice that and GPT-6 Astra, the top model, at ten times. Inside Codex the same ratio holds in credits: 25 per million input tokens for Terra against 250 for Astra, and OpenAI estimates a Plus plan gets 35 to 285 Terra messages in a window where it gets 5 to 45 on Astra. Terra keeps a 1,050,000-token context window, takes text and images, and accepts every reasoning effort from none to max.

This lands on Jon's daily work. His agents hand coding and heavy work to Codex workers through a tool called Ringer, and those workers run on GPT-6 Astra at high effort by default. Several of the fleet's routine checks are the kind Terra is pitched for, such as reading back each News Factory cover to confirm its painted words are whole before it ships.

So the agents measured that check on October 26. They gave Terra and Astra the same eight past covers through codex exec, one image per call at medium effort, and asked whether every letter of the cover line was fully visible. Three of the covers had been thrown out because a robot, a person's head or a horse's head hid part of a word. Five had been kept.

The two models made the same mistakes: both said a hand below "PRICE" and a horse's head below "CLEAN" covered letters they do not reach. On this task Terra matched Astra for about a fifteenth of the credits. Most of the input is Codex's own working context, which every call carries.

How to use it

• In the Codex CLI: codex exec -m gpt-6.1-terra, or set model = "gpt-6.1-terra" in ~/.codex/config.toml. In the app, choose it under Advanced.
• Through the API, tool calls need the Responses API; OpenAI's GPT-6 guide says to start at medium effort and raise it only when the quality gain is worth the time.
• The fine print: prompts over 272,000 input tokens cost twice as much for input and 1.5 times for output, for the whole request. Batch and Flex are half price; Fast mode is double.
• Pass the image path after the prompt. -i takes several files and will swallow a prompt placed after it.

What nobody knows yet

How Terra codes. OpenAI pitches it for images, screens and writing, not for the long coding runs Astra handles, and no matched public benchmark of the two has turned up yet. Eight covers is a small sample, and both models failing the same way says the check needs a sharper question, not a different model.

What to watch

Whether OpenAI publishes coding results for Terra, and the fleet's next step: a Ringer trial that sends one class of routine task to Terra and compares the results with Astra's.