GPT-6.1 Terra Launches for Work, Codex and the API
GPT-6.1 Terra adds a lower-priced option for visual reasoning, clear writing and computer use in ChatGPT Work, Codex and the API.
GPT-6.1 Terra launched on October 26 as OpenAI’s new option for visual reasoning, computer use and clear writing in ChatGPT Work, Codex and the API. Its Standard API input and output prices are half GPT-6.1 Sol’s. The release gives teams another model to evaluate for recurring professional work, particularly tasks that combine screenshots, documents and actions in an application. The exact API identifier is gpt-6.1-terra.
What GPT-6.1 Terra changes for model selection
OpenAI positions Terra around the work between reading information and acting on it: interpreting a technical drawing, matching a product photo, drafting a proposal, or entering information into an application. Its launch announcement reports improvements over GPT-5.6 Terra across professional computer tasks and writing. Those are OpenAI’s evaluations; teams should use their own recurring tasks to decide where Terra belongs.
For an existing Sol workflow, the immediate opportunity is a controlled comparison. Keep the instructions, input files and validation steps fixed, then change the model. A lower token price becomes useful when the completed result still meets the workflow’s requirements. If the cheaper run needs several corrections, compare the full conversation and review time rather than the first response alone.
The GPT-6.1 overview links the current models. The release does not rename original Sol, Luna or Astra, and historical examples retain the model that produced them. Teams can keep an established model for difficult steps while evaluating Terra on a narrower part of the process.
Where users can select the new model
The official Work and Codex changelog lists Plus, Pro, Business, Enterprise and Edu access. In Work on the web or the desktop app, open Advanced and select Terra. Work also supports the model on mobile. Terra is not available in ordinary Chat and does not support UAE inference residency. Enterprise availability follows workspace access and administrator controls.
Codex users can select the model with the model picker or launch a session with codex -m gpt-6.1-terra. The October 29 CLI release adds Terra to bundled, Vertex AI and Amazon Bedrock catalogs. A catalog entry describes a selectable model; the account, provider and workspace still determine whether the request is permitted. The access guide walks through those checks.
Before changing a team default, document which product the team uses. A subscription conversation in Work, a signed-in Codex session and an API application have different billing and configuration. Testing in one surface is useful evidence for that surface; it does not establish every other product’s availability or usage cost.
Prices, context and tools to check before migrating
The official API model page lists Standard prices of $1 per million ordinary input tokens, $0.05 cached input, $1.25 default cache writes and $5 output. The default write price uses a 30-minute minimum lifetime; a 12-hour minimum costs $2.50 per million write tokens. Longer input can change the applicable rate, so request size belongs in a migration estimate.
Terra has a 1,050,000-token context window and a 128,000-token output maximum. Its API reasoning efforts range from none through max, with medium the default. Tool calling requires the Responses API. Supported tools must still be configured by the application: selecting Terra does not itself connect a browser, grant filesystem access or supply an app integration.
For a text-only example, 10,000 ordinary input tokens and 2,000 output tokens cost $0.02 at Standard rates. That is an arithmetic example, excluding image and tool charges. Enter representative input, output and cache counts into the pricing calculator before estimating a monthly budget. Model-specific settings and the request example are on the Terra reference.
A useful first evaluation for writing and visual work
For a writing workflow, choose an existing document with a clear audience and a known final version. Give both models the same source material and ask for the same deliverable. Review names, dates, numerical claims and required actions before comparing style. Record whether the reviewer had to restore any facts that disappeared during rewriting.
For visual work, use screenshots or diagrams whose labels and values you can independently read. Ask for references to visible evidence, then compare the answer with the source. Include a difficult crop or an unclear value in the sample. The useful question is whether the model identifies the uncertainty where it matters to the task, rather than confidently completing a plausible interpretation.
For computer use, define the starting application state and the expected result. An invoice-entry test might require a saved draft with the correct supplier, amount and due date. Inspect the actual record afterward, including whether any unrequested changes occurred. Measure successful completion, corrections and elapsed time alongside token usage. This evaluation procedure is a suggested adoption method, not a reported Terra benchmark.
Keep a small set of representative tasks after the initial trial. A repeatable sample makes it easier to notice whether later prompt edits, tool changes or model selections affect the outcome. Deploy the model where that evidence supports it, then expand the scope as the workflow proves useful.