GPT-6.1 Terra Brings Screen Actions Into Its Visual Toolkit

The short version

• GPT-6.1 Terra began rolling out in ChatGPT Work and Codex on October 26, 2026.
• It focuses on interpreting images, operating computer interfaces and writing.
• Standard API pricing is $1 per million input tokens and $5 per million output tokens, half GPT-6.1 Sol’s rates.
• Codex CLI users need version 0.159.0 or later to select it.

OpenAI released GPT-6.1 Terra on October 26, 2026, giving eligible ChatGPT Work and Codex subscribers a model built for sharper visual reasoning, precise computer interaction and natural writing. The rollout covers the subscription plans listed below. Developers can also use it through the API, where standard input and output prices are half GPT-6.1 Sol’s rates and below those of its predecessor, GPT-5.6 Terra.

Who gets GPT-6.1 Terra in ChatGPT Work, Codex and the API?

GPT-6.1 Terra is rolling out to the following ChatGPT Work and Codex subscribers, while developers can access it through the OpenAI API.

Access by product and plan

In OpenAI’s launch description, a given ChatGPT Work or Codex allowance handles more tasks with GPT-6.1 Terra than GPT-5.6 Terra, and substantially more than GPT-6.1 Sol or GPT-6 Astra. OpenAI also describes its standard API prices as 50% below GPT-5.6 Terra and GPT-6.1 Sol; the exact token rates appear in the comparisons below.

What is GPT-6.1 Terra good at?

GPT-6.1 Terra is designed to interpret visual information, take precise actions through computer interfaces and produce clear, natural prose.

OpenAI’s ChatGPT account gives these examples:

• Decoding technical drawings.
• Drafting proposals.
• Filling forms.
• Editing videos.

The API documentation identifies additional uses:

• Finding the catalog entry that matches a product photograph.
• Interpreting technical drawings.
• Taking actions on application screens.
• Drafting text and revising existing writing.

OpenAI’s ChatGPT Work and Codex model guide also recommends Terra for these tasks:

• Interpreting charts.
• Reviewing visual materials.
• Completing forms using information in source documents.

How does GPT-6.1 Terra perform on OpenAI’s benchmarks?

OpenAI’s results show higher scores and lower task costs than GPT-5.6 Terra on its professional computer-task benchmark, with comparisons against the larger GPT-6 models varying by effort setting.

Agents’ Last Exam v1 evaluates professional computer work. Its tasks include producing a 3D model from architectural drawings and creating a restoration log from an animated film. All scores and task costs here come from OpenAI.

Agents’ Last Exam scores and costs

OpenAI describes Terra’s medium-effort result as a 15-percentage-point improvement over GPT-5.6 Terra for roughly a third of the cost. It matches GPT-6 Astra’s medium-effort score while charging one-tenth as much per token.

At max effort, OpenAI puts Terra’s task cost at roughly nine times lower than Claude Opus 5.5’s. Its launch measurements also show task completion taking less than a third of the time required by GPT-6 Astra and GPT-6.1 Sol.

OpenAI reports these additional results:

• MMMU-Pro, which measures visual reasoning: Terra exceeds GPT-5.6 Terra at every effort setting and approaches GPT-6 Astra and GPT-6.1 Sol while costing substantially less.
• OSWorld 2.1, which tests desktop application tasks: medium-effort Terra exceeds GPT-5.6 Terra’s highest score while using a fraction of its task cost.
• Error-inducing prompts, an internal OpenAI evaluation: low-effort errors fall from 53.6% for GPT-5.6 Terra to 16.7% for GPT-6.1 Terra. At max effort, Terra reaches 4.3%, an 18-percentage-point reduction that puts it near GPT-6 Astra and GPT-6.1 Sol.
• Terminal-Bench 4.0, which evaluates command-line coding: Terra scores 50.3% at max effort, compared with GPT-5.6 Terra’s published 21.5%.
• Terminal-Bench-Science 0.1, which evaluates scientific work performed through code: max-effort Terra exceeds GPT-6.1 Sol at lower cost and finishes within five percentage points of Claude Opus 5.5 for less than one-fifth its estimated cost.

OpenAI still recommends GPT-6.1 Sol or GPT-6 Astra for the hardest coding work, and its Developer Community announcement recommends those models for in-depth research as well.

Safety protections and evaluation results

OpenAI classifies GPT-6.1 Terra at the Critical cybersecurity threshold in its Preparedness Framework, with the protections developed for GPT-6 Astra. Its evaluations found better adherence to task scope and more disclosure of limitations and blockers than GPT-5.6 Terra. OpenAI also reports strong resistance during its prompt-injection testing; the GPT-6.1 Terra system card details the safeguards and evaluations.

We ran it: how did GPT-6.1 Terra perform in nerdschalk’s Codex tests?

GPT-6.1 Terra returned correct results in all nine runs and recorded the lowest median time for each of our three tasks.

We ran the comparison with this setup:

• Machine: a Linux server.
• Dates: the night of October 26, 2026, to October 27, 2026, minutes after the launch.
• Tool: Codex CLI 0.159.0, signed in with ChatGPT.
• Reasoning effort: medium for all three models.
• Repetitions: three runs per model for each task.

We made these tasks so each had one exact correct answer; the drawing and form tasks also match examples OpenAI gives for Terra:

• A Python code fix checked against eight tests.
• Our mounting-plate technical drawing, with ten questions including two calculation questions.
• A supplier form with 15 fields, completed using an email.

Correct results and measured completion times

Terra’s median total was 27.6 seconds, about 45% below Sol’s 49.7 seconds at the same effort. Sol also produced correct results in every run.

OpenAI reports that Terra completes tasks over three times faster than Sol and Astra on its benchmark; on our three short medium-effort tasks, Sol took about 1.8 times as long as Terra, a smaller gap in the same direction.

Luna returned an empty form in two of its three attempts. It said it would read the email file but never opened it. In one drawing attempt, it also missed both calculated answers while correctly reading the other eight answers.

All three models fixed the code without modifying its tests. Each passed all eight tests in every code run.

This was a small comparison with these limits:

• Three short tasks we created, with three runs each.
• One machine and one night of testing.
• Medium reasoning effort throughout.
• Our measured times include Codex overhead and depend on OpenAI’s server conditions at the time.
• Computer use, video editing and long writing were not tested in this round.

How does GPT-6.1 Terra compare with GPT-5.6 Terra, Sol, Luna and Astra?

GPT-6.1 Terra lowers the previous Terra’s token prices, extends its Codex context window and sits between Sol and Luna in standard API pricing.

The GPT-5.6 Terra model page describes the older model as balancing capability and cost, roughly equivalent to the mini tier in earlier GPT-5 families.

New Terra versus the previous model

The newer Terra has a later knowledge cutoff, cheaper input and output, and more room for context in Codex. Its API context and maximum output remain the same as the older model’s.

OpenAI’s GPT-6 guide recommends comparing the two Terras on your own images and workflows at matching reasoning effort, and moving Chat Completions requests that call tools to Responses.

GPT-6 model labels and standard prices

OpenAI’s Work and Codex guidance assigns the models these roles:

• GPT-6 Astra: the most demanding work.
• GPT-6.1 Sol: capability approaching Astra at a lower price.
• GPT-6.1 Terra: everyday writing, visual interpretation and tasks across supported applications.
• GPT-6 Luna: focused, repeatable work at high volume, including extraction, classification, transformation and structured summaries.

The guide gives Terra the same capability rating as Astra and Sol, while its speed rating exceeds both and matches Luna.

OpenAI gives this example prompt for Terra: “Review these dashboard screenshots and summarize the main changes. Refer to the chart labels and values, and flag anything you can’t read confidently.” It also recommends checking image-derived names, numbers and other details against the original material.

Terra’s API supports none reasoning effort. Astra and GPT-6.1 Sol do not.

These API pricing adjustments apply to GPT-6.1 Terra:

• Cached input costs 5% of the ordinary input rate.
• A request exceeding 272K input tokens incurs double input and cache rates, plus 1.5 times the output rate, across the entire request.
• Batch and Flex processing each cost 50% below Standard.
• Fast mode costs twice the Standard rate.
• Cache writes cost $1.25 per million tokens with 30-minute retention, or $2.50 with 12-hour retention, before long-context adjustments.
• Regional processing carries a 10% premium where offered.

How do you pick GPT-6.1 Terra in ChatGPT Work or Codex?

Use the model control in your client, or launch Codex CLI with gpt-6.1-terra.

• ChatGPT Work on the web: choose the model using the model and reasoning control beneath the composer.
• ChatGPT desktop app: use the control beneath the composer, open Advanced, then choose GPT-6.1 Terra from the model list.
• Codex CLI: install version 0.159.0 or later. Launch it with the command below, or enter /model during a session to change the model and effort.
• Codex IDE extension: choose it using the model switcher beneath the composer.

To start Codex CLI with Terra, enter this command in your terminal:

For a persistent default, set model = "gpt-6.1-terra" in config.toml, using the default-model setting described in OpenAI’s Codex configuration documentation.

Codex catalog defaults and Fast labels

Those settings come from OpenAI’s model catalog downloaded by Codex CLI 0.159.0 on October 26, 2026. Terra’s Codex reasoning choices are:

• low
• medium
• high
• xhigh
• max
• ultra

Codex describes ultra as “Maximum reasoning with automatic task delegation.” The model guide explains that Ultra uses subagents for larger jobs that can be divided into separate tasks.

What does GPT-6.1 Terra support for API developers?

GPT-6.1 Terra supports Responses, Chat Completions and Batch, with text and image input and text output; tool calls require Responses.

The API does not support audio or video input for this model. Fine-tuning and predicted outputs are also unsupported.

These tools are supported through Responses:

• Web search.
• File search.
• Image generation.
• Code interpreter.
• Hosted shell.
• Apply patch.
• Skills.
• Computer use.
• MCP.
• Tool search.

Streaming, function calling and structured outputs are supported. Fast mode also supports EU data residency for GPT-6.1 Terra.

Standard API rate and queue limits

Your API usage tier determines the applicable request, token and Batch queue limits.

FAQs

Is GPT-6.1 Terra free or available on the API Free tier?

The API Free tier does not support GPT-6.1 Terra. OpenAI names these subscription plans for its ChatGPT Work and Codex rollout:

• Plus
• Pro
• Enterprise
• Edu

Is GPT-6.1 Terra in the API?

Yes. It was released in the API on October 26, 2026, with model ID gpt-6.1-terra. Responses, Chat Completions and Batch are supported; tool calling requires Responses.

How much does GPT-6.1 Terra cost in the API?

Standard pricing per million tokens is $1 for input, $0.05 for cached input and $5 for output. Batch and Flex cost 50% less than Standard, while Fast mode costs twice as much. Requests exceeding 272K input tokens receive long-context pricing across the full request.

GPT-6.1 Terra vs GPT-6.1 Sol: which is faster and cheaper?

Terra’s standard API input and output rates are half Sol’s: $1 versus $2 for input, and $5 versus $10 for output, per million tokens. OpenAI’s model guide gives Terra the higher speed rating. In nerdschalk’s three-task Codex comparison at medium effort, the median totals were 27.6 seconds for Terra and 49.7 seconds for Sol.

How do I select GPT-6.1 Terra in Codex CLI?

1. Use Codex CLI version 0.159.0 or later.
2. Launch with codex -m gpt-6.1-terra, or enter /model in an existing session and choose GPT-6.1 Terra.
3. To make it the default, set model = "gpt-6.1-terra" in config.toml.

What is GPT-6.1 Terra’s knowledge cutoff?

GPT-6.1 Terra’s listed knowledge cutoff is June 1, 2026. GPT-5.6 Terra’s cutoff is February 16, 2026.

How large is GPT-6.1 Terra’s API context window?

The API context window is 1,050,000 tokens, with a maximum input of 922,000 tokens and maximum output of 128,000 tokens. Codex lists a separate 500,000-token context window for GPT-6.1 Terra.

What safety protections does GPT-6.1 Terra have?

GPT-6.1 Terra uses the safeguards stack applied to GPT-6.1 Sol and GPT-6 Astra. OpenAI classifies it as Critical for cybersecurity and High for biological and chemical capability under its Preparedness Framework. Its evaluations report strong resistance to prompt injection, alongside improved scope adherence and disclosure of blockers compared with GPT-5.6 Terra.

---

Related on NerdsChalk

• ChatGPT Dots Gain Phone Setup and Codex Tasks
• Codex on Windows Gets an MXC Sandbox With No Admin Setup: Check If Your PC Supports It
• Codex New Threads Not Working in ChatGPT Desktop? Check the Affected Version
• Codex Auto-Review Is Now Free for ChatGPT Sign-Ins