GPT-6.1 Terra API

GPT-6.1 Terra is OpenAI's new second tier from the bottom, and it is built for a specific kind of work: reading screens, clicking through apps, filling in forms, pulling numbers out of documents and writing the email or proposal that comes after. Released on October 26, 2026, it slots between GPT-6 Luna and GPT-6.1 Sol, and OpenAI introduces it in one line: fast and capable computer use for the whole workday.

The headline is speed at near-flagship skill on that kind of task. On Agents' Last Exam, OpenAI's computer-use benchmark, GPT-6.1 Terra at medium effort scores 57.6%, 15 points above GPT-5.6 Terra and on par with GPT-6 Astra. OpenAI says it finishes those tasks more than three times faster than Sol or Astra. On OSWorld 2.1, which has agents operating desktop apps through multistep jobs, it reaches 74.2% at maximum effort, a few points below the larger models.

It also hallucinates far less than the Terra it replaces. On OpenAI's set of prompts designed to provoke factual mistakes, the share of answers containing an error falls from 53.6% to 16.7% at low effort and from 22.3% to 4.3% at max effort, nearly level with Sol and Astra. For an agent that types into a form or a model that drafts client documents, fewer invented facts matter as much as raw intelligence.

Where Terra sits in the GPT-6 family

OpenAI now names its models by class rather than by size. From lowest to highest the four classes are Luna, Terra, Sol and Astra. GPT-6 Astra opened the GPT-6 generation in early September, GPT-6 Sol and GPT-6 Luna followed on September 22, and GPT-6.1 Sol arrived a week later. Until now the GPT-6 generation had no Terra, so teams that liked GPT-5.6 Terra from July had to choose between a smaller Luna and a heavier Sol. GPT-6.1 Terra fills that gap and joins Sol in the 6.1 family.

OpenAI's system card describes it as "a fast, efficient model for high-volume professional work, with strong image analysis and computer use capabilities." The marketing frames the same idea more loosely: a model that understands what is on screen, takes action across apps and writes naturally. OpenAI is also clear about what it is not. For in-depth research or complex coding, the company still points people to GPT-6.1 Sol or GPT-6 Astra.

Like every GPT-6 model, Terra is closed. There are no downloadable weights and no published parameter count, and use is governed by OpenAI's usage policies and service terms. Developers reach it through the OpenAI API as gpt-6.1-terra, through GitHub Copilot and Amazon Bedrock, and through resellers including Unicorn API. In OpenAI's own apps it is available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu accounts, but not in the standard ChatGPT chat window.

What changed from GPT-5.6 Terra

Computer use. This is the biggest jump. Agents' Last Exam at medium effort rises from 42.6% to 57.6%. On OSWorld 2.1, Terra at medium effort scores 66.7%, already above GPT-5.6 Terra's best result at any effort level, and 74.2% at max. An OpenAI staffer reported that Terra filled in one web form more than two minutes faster than Sol did, and summed it up as "similar on vision and computer use (but much faster!)" compared with Astra and Sol.

Visual reasoning. On the MMMU-Pro multimodal test, Terra scores 84.6% at medium effort against 86.3% for GPT-6.1 Sol, and up to 85.5% at higher settings. OpenAI places it above GPT-5.6 Terra and close to Astra and Sol at every effort level. That matters for agents, because computer use starts with reading a screenshot correctly.

Coding. Terminal-Bench 4.0, which measures agents working in a command line, more than doubles: 50.3% at max effort, against 21.5% for GPT-5.6 Terra. On Terminal-Bench-Science 0.1, which focuses on research-style coding, Terra scores 58.3% at max, slightly above GPT-6.1 Sol and within five points of Claude Opus 5.5.

Factuality. The error rates above are the clearest step. At low effort the old Terra got more than half of the hard prompts wrong; the new one gets about one in six wrong. The system card adds that Terra makes substantially fewer factual errors than GPT-6 Sol and is comparable to GPT-6.1 Sol when reproducing cases users had flagged as hallucinations.

Writing. OpenAI pitches clearer, more natural prose, aimed at proposals, reports and messages rather than chat banter.

A true no-reasoning mode. Terra accepts none as a reasoning effort, alongside low, medium, high, xhigh and max. GPT-6.1 Sol always thinks; Terra can answer immediately when the task is simple.

What OpenAI says about how it is built

OpenAI publishes little about the model's internals. The system card says GPT-6.1 Terra uses the same types of data and training as GPT-6 Astra, without describing a new method, and gives no size or architecture details.

The safety picture is detailed. Under OpenAI's Preparedness Framework, Terra is treated as Critical capability in cybersecurity and High in biology and chemistry, and below High on AI self-improvement, so it ships with the same class of safeguards as the larger models. In instruction-hierarchy tests, defenders succeeded against about 99.99% of direct attacks, and OpenAI calls its robustness to indirect prompt injection, the kind hidden in web pages and documents an agent reads, comparable to Astra and Sol. In alignment tests it made zero attempts to bypass automatic review, OpenAI found no hidden reasoning in its chain of thought, and its rate of deceiving users about blockers is in line with recent models and much better than GPT-5.6 Terra.

Health results are mixed. MentalHealthBench rises from 46.2% to 55.5%, and HealthBench Professional gains 5.0 points, but the broader HealthBench drops 1.9 points and HealthBench Hard 1.3. The card also records small, statistically non-significant regressions on vision tests for hateful and extremist content.

GPT-6.1 Terra specs

Artificial Analysis measured about 91 output tokens per second, faster than GPT-6.1 Sol's roughly 74 but below the roughly 132 it recorded for GPT-5.6 Terra.

Where GPT-6.1 Terra is strong

Agents that work a screen. Browser and desktop agents spend most of their time looking at screenshots and deciding where to click. Terra matches the larger models closely on exactly those skills and gets through each step faster, so a twenty-step task finishes sooner and a human waiting on it gets the result while it is still useful.

Forms and data entry. Filling in web forms, copying fields between systems and checking entries against a source document are the jobs OpenAI names first. Low factual error rates make it safer to let it type.

Documents and extraction. In one hands-on review, Terra extracted every field from an invoice correctly, including supplier, invoice number, dates, currency and tax lines, and its arithmetic checked out. In a video review task it kept all eight sampled intervals in order and marked the gaps that needed closer inspection rather than inventing details it could not see.

Writing that keeps its promises. In the same review, a client proposal kept every commitment from the brief: the exclusions, two review rounds, client-supplied photos and copy, and an open question about which five screens were in scope. The prose was described as clear and usable.

Research-style command-line work. Its Terminal-Bench-Science result, slightly above GPT-6.1 Sol, shows it can handle data wrangling and scientific scripting better than its tier suggests.

Huge context. The 1,050,000-token window is the same as Sol's, so a long manual, a stack of contracts or a full session of agent screenshots fits in one request.

Weak spots and trade-offs

The default effort undersells it. Artificial Analysis scores Terra 51 on its Intelligence Index v5.0 at max effort, but 45 at xhigh, 42 at high, 38 at medium and 29 at low. Medium is the API default, and at that setting it sits much closer to GPT-6 Luna's 28 than to Sol's 54.5. For anything beyond routine work, raise the effort.

Coding is not its job. On Terminal-Bench 4.0 it trails GPT-6.1 Sol's 55.9% and is far behind Claude Sonnet 5.5 at 70.6%. On Artificial Analysis's coding agent index it scores 48 against Sol's 54. OpenAI's staffer called it "somewhat weaker on most other tasks" outside vision and computer use.

Verbose at the top setting. At max effort it writes about 55% more output tokens than Sol, which eats into its speed advantage on long answers and fills the 128,000-token output limit sooner.

Plausible but wrong. The hands-on review caught it misreading a technical drawing: it took a 5-unit horizontal setback as a vertical height difference and produced a 3D mesh that was valid but wrong. The same review found a proposal of 217 words where 230 to 300 were requested. Outputs can look right and pass a format check while missing a detail, so verify what matters.

Positioning debate. The Hacker News launch thread reached 453 points and 285 comments, and much of it asked why Terra needs to exist when you could use Sol at a lower effort. One commenter called it "a weird middle child." Builders running agent workloads were more positive.

Text out only. It reads images but does not generate images or audio.

How to prompt GPT-6.1 Terra

Give it the target, the source and the stopping rule. Terra does best on concrete, checkable work, so describe the end state rather than the process, and ask it to confirm what it entered. Use none or low for quick classification and rewrites, medium for routine agent steps, and high or above for anything that involves judgement or long documents.

For a form-filling agent:

Using the open browser, enter the supplier details from the attached invoice into the vendor form. Leave a field empty if the invoice does not show it. Before submitting, list every field you filled with its value, and stop if the form asks for anything not on the invoice.

For extraction:

Read this invoice image and return JSON with keys supplier, invoice_number, issue_date, currency, subtotal, tax and total. Recompute the total from the line items and add a key total_matches set to true or false.

For writing, state hard limits as numbers and repeat them at the end:

Draft a project proposal of 230 to 300 words from the notes below. Keep every exclusion and the two review rounds exactly as written. End with the open question about scope. Count the words before you finish and adjust if you are outside the range.

When a task depends on spatial or numeric detail in a drawing or chart, ask it to describe what each label refers to before acting on it.

GPT-6.1 Terra versus Sol, Astra, Luna and Claude

GPT-6.1 Sol is the stronger all-rounder, with clear leads in coding and general reasoning. Pick Sol for software engineering and hard analysis; pick Terra when the work is mostly screens, forms and documents and turnaround time matters.

GPT-6 Astra remains OpenAI's most capable model. Terra matches it on Agents' Last Exam at medium effort while running much faster, so Astra is worth it only where its depth on hard problems is needed.

GPT-6 Luna is the small tier for simple, high-volume calls. Terra at xhigh scores 45 on Artificial Analysis's index against Luna's 28 at max, so move up to Terra when Luna's answers are not good enough.

GPT-5.6 Terra is faster per token, but the new model wins on every benchmark OpenAI reported.

Claude Sonnet 5.5 and Claude Opus 5.5 score higher on Artificial Analysis's index and, in Sonnet's case, far higher on Terminal-Bench 4.0. For coding agents, test them first; for screen-driving agents and office documents, Terra is the OpenAI model to compare against them.

Getting started

Try a prompt in the Playground, then create a key on the API keys page. On Unicorn API the model answers on the OpenAI-compatible /v1/chat/completions endpoint with model: "gpt-6-1-terra". Images go in as image_url parts, function tools use the standard tools array, and response_format returns JSON. Set reasoning_effort to the level you need, or leave it out for the medium default, and give max_tokens room for thinking, since reasoning tokens count toward it. The docs cover streaming, tools and errors.

Verdict

GPT-6.1 Terra is a specialist with a clear brief. On computer use and visual reasoning it comes close to OpenAI's biggest models and gets there several times faster, and it cuts factual errors sharply compared with GPT-5.6 Terra. It is not the model for hard coding, its default effort leaves a lot of quality on the table, and it can produce confident output that misses a detail. For browser agents, form filling, document extraction and the writing that surrounds them, it is the GPT-6 model built for the job.