GPT-6.1 Terra cuts token prices; tool calls require Responses

A cheaper model for visual work

OpenAI released GPT-6.1 Terra on October 26 for visual reasoning, writing and computer-based tasks. Its standard API input and output prices are half those of GPT-5.6 Terra and GPT-6.1 Sol, according to the launch announcement. The company is also rolling it out through ChatGPT Work and Codex.

For developers, the change extends beyond price. Terra supports tool calls through the Responses API, so an application that currently calls tools through Chat Completions cannot migrate simply by changing its model name. That distinction matters for the screen-reading and application-control workflows OpenAI is promoting.

Source 1Source 2Source 3Source 4

What an API migration involves

OpenAI’s migration guide specifically directs GPT-5.6 Terra applications using Chat Completions tool calls to move those requests to Responses. This is a requirement for that integration pattern, not a claim that Chat Completions has disappeared: the dated release entry lists both endpoints.

The model accepts text and images and produces text. Its published capacity is a 1,050,000-token context window, with a separate maximum of 128,000 output tokens; native audio and video are not supported. OpenAI offers reasoning settings from none through max, with medium as the default. Its guide recommends comparing the old and new Terra at the same effort on representative images and workflows.

The API retirement schedule reviewed for this report does not list a GPT-5.6 Terra shutdown. This launch therefore supplies no retirement deadline that would, by itself, force existing users to switch.

Source 2Source 3Source 4Source 9

The headline rates are only part of the bill

Standard pricing is $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens. For a hypothetical request using 100,000 uncached input tokens and 10,000 billable output tokens, those base rates total $0.15. That calculation excludes cache writes, tool or hosting charges and processing premiums; it is not a measured task cost.

The model documentation adds a consequential threshold: above 272,000 input tokens, input and cache rates double and output rates rise by 50% for the entire request. Having room for a large document collection does not mean every request uses the advertised base prices.

Cache writes also cost more than cache reads: standard writes are $1.25 per million tokens for 30-minute retention or $2.50 for 12 hours. Fast mode doubles standard rates; regional processing adds 10% where available. These conditions make a blanket promise of halving a workflow’s cost premature.

Source 3Source 7

Work and Codex access differs from ordinary Chat

OpenAI’s current release notes list a rollout to Plus, Pro, Business, Enterprise and Edu in ChatGPT Work and Codex. That includes Business, which the launch announcement’s availability paragraph does not name. Availability depends on the account, region and workspace settings as the rollout reaches each user; Terra is not available in regular Chat.

The release notes also specify that Terra does not support UAE inference residency. This concerns where inference runs, rather than the user’s physical location. In Work and Codex, availability still depends on the account and workspace; Enterprise access requires the relevant administrator controls. A published plan list does not establish that a particular organization has enabled the model.

API usage is billed separately from ChatGPT subscriptions. Developers using the command-line client need Codex CLI 0.159.0 or later. We have not checked a customer account or verified its model availability. Those distinctions matter before a team changes a shared workflow or assumes its existing subscription covers API requests.

Source 1Source 5Source 8Source 11Source 12

Reading a screen still requires an application to act

Consider a hypothetical catalog-maintenance task: compare a product photograph with an existing listing, then prepare a description update for review. OpenAI names photo-to-catalog matching as a Terra use case. The distinction between recognizing the mismatch and changing the listing is the integration work: the application executes tools, keeps the browser or desktop session available and returns observations, including screenshots.

OpenAI’s computer-use guide treats those screenshots as billable image input and says returned screen content goes to OpenAI. It also warns that setting store to false does not eliminate every form of retention. For a listing editor, our reading is that the useful deliverable is a verifiable proposed change with a controlled save step. The guide calls for restricted environments, approval for consequential actions and checking the actual saved result.

Source 3Source 6

Performance claims remain OpenAI’s results

OpenAI reports a 50.3% result for Terra at maximum reasoning effort on Terminal-Bench 4.0, compared with GPT-5.6 Terra’s published 21.5%. It nevertheless recommends GPT-6.1 Sol or GPT-6 Astra for the most complex coding work. These figures describe the company’s evaluation, not our testing or a success rate for the catalog example.

The system card reports a mixed safety picture from an internal Codex deployment simulation. At severity levels 1 or 2, Terra had higher flagged rates in some categories, including destructive action and inbound prompt injection, than GPT-6.1 Sol and GPT-6 Astra. That does not mean its more severe aggregate was higher: across 48,690 matched tasks, Terra received 22 severity-3-or-higher flags, versus 24 for Sol and 26 for Astra. OpenAI cautions that this simulation is a signal of internal deployment risk, not a direct measure of external safety. Lower token prices do not resolve that distinction or establish a production workflow’s reliability.

Source 1Source 10