GPT-6.1 Terra launches: access, pricing and what to test before automating screen work

What launched on October 26?

OpenAI’s API changelog records the release of GPT-6.1 Terra on October 26, 2026. The company positions it for visual reasoning, computer use and natural writing.

The launch post lists availability in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, alongside API access under gpt-6.1-terra. Product access and developer integration are separate routes: a model appearing in an application does not configure an API workflow for you.

Public record: OpenAI — GPT-6.1 Terra launch and product availability · OpenAI API changelog — October 26 release

The computer-use claim has a specific evaluation behind it

On Agents’ Last Exam, OpenAI reports that Terra at medium reasoning effort scores 15 percentage points above GPT-5.6 Terra at about one-third the cost. The launch post also reports task completion more than three times faster than GPT-6 Astra and GPT-6.1 Sol. These are vendor evaluation results; Promptwire has not independently reproduced them.

The practical implication is an evaluation opportunity. A team that repeatedly moves information from documents into software can test that exact route. A benchmark improvement alone does not establish whether an unusual invoice layout, a duplicate record or a changed screen will be handled correctly.

Public record: OpenAI — GPT-6.1 Terra launch and product availability

API integration: tools require Responses

The model page states that tool calling requires the Responses API. It lists text and image input with text output, a 1,050,000-token context window, 128,000 maximum output tokens and a June 1, 2026 knowledge cutoff.

Reasoning effort options are none, low, medium, high, xhigh and max, with medium the default. These settings belong in your test record. When comparing runs, changing the model and the effort setting together makes it harder to identify what improved.

The model page lists support for computer use, function calling and structured outputs. Those capabilities still need an application environment, tools and a definition of success. Inspect the saved record as well as the assistant’s account of what it did.

Public record: OpenAI API — Terra specifications, tools and pricing conditions

Standard rates are only the starting point

OpenAI’s standard short-context API table lists the following US-dollar rates per million tokens. These are API usage charges, rather than ChatGPT subscription prices.

Public record: OpenAI API — current pricing tables

Long requests change the calculation

The model documentation says prompts above 272K input tokens use twice the input and cache rates and 1.5 times the output rate for the full request. Fast mode is priced at twice Standard; Batch and Flex are 50% below Standard. Regional processing adds 10% where available.

For budgeting, record service tier and context length alongside token totals. Also account for the tools your workflow uses. A low token price does not by itself tell you what a completed multi-step job will cost.

Public record: OpenAI API — Terra specifications, tools and pricing conditions

A useful pilot: reconcile a document with a saved draft

Promptwire’s suggested trial uses fictional invoices and a test accounting workspace. Include one straightforward invoice, one with an unclear date, one with a credit adjustment and one whose invoice number already exists. Define the correct draft record for each before running the agent.

Ask it to identify the supplier, invoice number, currency, line items and total, then prepare a draft entry. Require a clarification when a necessary detail is unreadable or ambiguous. Give the workflow a clear stopping point: a draft ready for review, with no payment or external submission.

Check the application after the run. Did the invoice number survive the transfer? Were currency and decimal values preserved? Was the duplicate flagged rather than entered again? Did the report accurately say what was saved and what still needed review?

Compare several runs with your current process. Record accepted drafts, corrections, time spent reviewing, elapsed time and total usage charges. A model that completes a screen sequence quickly may still create work for a person if the destination fields are wrong.

This is a proposed evaluation, not a report of testing Terra. The launch offers a reason to run a focused pilot; the evidence for adoption should come from the accuracy of the final records and the effort needed to approve them.