What Is GPT-6.1 Terra? Pricing and Benchmarks

Where GPT-6.1 Terra fits in the GPT-6 lineup

GPT-6.1 Terra is the third of four GPT-6 models by API price, sitting below GPT-6 Astra and GPT-6.1 Sol and above GPT-6 Luna. OpenAI shipped it on October 26, 2026, as the successor to GPT-5.6 Terra (see GPT-5.6 Sol, Terra and Luna: Specs and Pricing), and calls it a big jump over that model. OpenAI pitches three strengths:

• Reading images in fine detail
• Accurate computer use, where the model looks at a screen and clicks or types its way through an app
• Clearer, more natural writing

In the API it takes up to 1,050,000 tokens of context and accepts text and images as input. Output is text only.

The four GPT-6 models at a glance

Based on OpenAI's launch post and API pricing page as of October 29, 2026.

GPT-6 Astra

Flagship. OpenAI's pick, with Sol, for the hardest coding

GPT-6.1 Sol

Upper tier. Twice Terra's price

GPT-6.1 Terra

Mid tier. Built for computer use and high-volume routine work

GPT-6 Luna

Cheapest model in the API

Built for drawings, forms and everyday computer tasks

OpenAI's own examples of what Terra does well are drafting proposals, reading technical drawings, filling in forms and editing video. These are the writing and on-screen chores that eat most of a working day.

For high-volume work, OpenAI points to a job like pulling line items off incoming invoices and keying them into accounting software. The lower price is what makes that kind of automation pay off at scale, according to OpenAI.

Available in ChatGPT Work, Codex and the API

From launch day, every Plus, Pro, Enterprise and Edu user can pick GPT-6.1 Terra in ChatGPT Work and Codex. Developers call it in the OpenAI API with the model name gpt-6.1-terra. The launch post doesn't list the free plan or Business.

In ChatGPT Work and Codex, the same usage allowance stretches to more tasks on Terra than on GPT-5.6 Terra, and to far more than on GPT-6.1 Sol or GPT-6 Astra. If you keep running into limits on Sol, moving routine tasks to Terra frees up room.

GPT-6.1 Terra pricing is half of Sol

On the API, Terra charges $1 for every million input tokens and $5 for every million output tokens. A token is the small chunk of text a model reads and writes, and you pay per million of them. Those rates are exactly half of GPT-6.1 Sol's and a tenth of GPT-6 Astra's.

Cached input, meaning text the model has already seen and can reuse, costs $0.05 per million. Compared with GPT-5.6 Terra, which charged $2 for input and $12 for output, input is half the price and output is about 58% cheaper.

API prices next to Astra, Sol and Luna

Among the four GPT-6 models and the previous Terra, only Luna is cheaper than GPT-6.1 Terra. Here are the standard input and output rates side by side.

API prices for GPT-6.1 Terra and other OpenAI models

From OpenAI's API pricing and model pages as of October 29, 2026. Standard rates in US dollars per million tokens. Bar height is relative to the highest value.

Input

$10GPT-6 Astra

$2GPT-6.1 Sol

$2GPT-5.6 Terra

$1GPT-6.1 Terra

$0.10GPT-6 Luna

02.557.510$

Output

$50GPT-6 Astra

$12GPT-5.6 Terra

$10GPT-6.1 Sol

$5GPT-6.1 Terra

$0.50GPT-6 Luna

02040$

Per-call numbers make the gap easier to feel. Feed a model a 100K-token document and have it write 10K tokens back, and one call costs this much:

Cost of one call: 100K tokens in, 10K tokens out

Calculated from OpenAI's standard API rates. US dollars. Bar height is relative to the highest value.

$00.511.5

$1.50GPT-6 Astra

$0.32GPT-5.6 Terra

$0.30GPT-6.1 Sol

$0.15GPT-6.1 Terra

Run that same call 1,000 times a day and Sol comes to $300 while Terra comes to $150. The more calls you make, the wider the gap gets.

Long prompts and Fast mode cost more

Once a single request goes past 272K input tokens, the whole request is billed at twice the input and cache rates and 1.5 times the output rate. If you feed it long documents, splitting them to stay under 272K avoids the surcharge.

Fast mode costs double the standard price in exchange for quicker responses. Pinning processing to a specific region adds 10%. Going the other way, Batch and Flex are half price, which suits bulk jobs that don't need an answer right away.

How much better is it than GPT-5.6 Terra?

GPT-6.1 Terra beats GPT-5.6 Terra on every test OpenAI published, covering computer use, factual accuracy and coding. It is also quick: OpenAI says it finishes tasks more than three times faster than GPT-6 Astra or GPT-6.1 Sol. Every number in this section comes from OpenAI's own evaluations.

GPT-5.6 Terra vs GPT-6.1 Terra on two OpenAI tests

OpenAI's own evaluations, published October 26, 2026. Bars are percentages on a 0 to 100 scale. Lower is better for error rate; higher is better for tasks solved.

Error rate on error-inducing prompts (low effort)

%02040

53.6%GPT-5.6 Terra

16.7%GPT-6.1 Terra

Terminal-Bench 4.0 tasks solved (max effort)

%02040

21.5%GPT-5.6 Terra

50.3%GPT-6.1 Terra

Astra-level computer use at a tenth of the price

On Agents' Last Exam, GPT-6.1 Terra scored 15 points above GPT-5.6 Terra while spending roughly a third as much. The benchmark measures computer use on professional work. Two of its tasks:

• Building a 3D model from architectural drawings
• Logging an animated film shot by shot, with timecodes, for a restoration team

That score came at medium reasoning effort. Reasoning effort is the setting for how long the model thinks before it answers, and the API offers six levels, from none up to max.

At the same medium setting, Terra's score matched GPT-6 Astra's, at one-tenth Astra's price per token. Turned up to max, it used tokens more than three times as efficiently as Anthropic's Claude Opus 5.5, at a quarter of Opus's per-token price. OpenAI estimates the resulting cost per task is roughly 9x lower.

Terra also beat GPT-5.6 Terra at every effort setting on MMMU-Pro, a test of visual reasoning, and came close to Astra and Sol for far less money. On OSWorld 2.1, which has the model work through multistep tasks in desktop apps, Terra at medium topped GPT-5.6 Terra's best score.

Error rate falls from 53.6% to 16.7%

On an internal OpenAI test built from prompts designed to trip models up, Terra's error rate at low effort dropped to 16.7%, down from 53.6% for GPT-5.6 Terra. That cut of about 37 points means it now gets roughly one answer in six wrong instead of one in two.

At max effort the error rate dropped 18 points to 4.3%, about the same as GPT-6 Astra and GPT-6.1 Sol. For work where facts have to be right, raise the effort level.

Coding scores more than double the old Terra

On Terminal-Bench 4.0, where coding agents fix problems from the command line, GPT-6.1 Terra at max effort solved 50.3% of tasks. GPT-5.6 Terra's published score was 21.5%, so the new model more than doubles it. One example task is repairing a web page whose layout jumps around while it loads.

On Terminal-Bench-Science 0.1, which has agents carry out research tasks in code, Terra outscored GPT-6.1 Sol while costing less. It finished within five points of Claude Opus 5.5 at under a fifth of the estimated cost.

Terra keeps Astra's safeguards

Under OpenAI's Preparedness Framework, GPT-6.1 Terra's cybersecurity capability reaches the Critical threshold, and the model ships with the protections OpenAI built for GPT-6 Astra. OpenAI's evaluations also found Terra less likely than GPT-5.6 Terra to wander outside what it was asked to do, and more likely to tell the user plainly about limits or anything blocking the task. It held up well against prompt injection, where instructions hidden in a web page or document try to pull an AI agent off its job.

Should you use GPT-6.1 Terra or Sol?

Send high-volume routine work and computer-use tasks to Terra, and keep the hardest coding on Sol or Astra; that split follows OpenAI's own guidance. Terra costs half as much as Sol, and on OpenAI's computer-use benchmark it matched Astra. Here is how the jobs map out.

Which GPT-6 model to pick for each kind of job

Based on the example tasks and the routing note in OpenAI's launch post, October 26, 2026.

Job	Model to use
Reading invoices and entering them into accounting software	GPT-6.1 Terra
Filling in forms and other app work	GPT-6.1 Terra
Drafting proposals, reading technical drawings	GPT-6.1 Terra
The most complex coding	GPT-6.1 Sol or GPT-6 Astra

High-volume and computer-use work suits Terra

Jobs like copying invoice line items into accounting software, filling in forms and drafting proposals are where Terra fits best. They are the examples OpenAI chose, and the lower the cost per item, the more it pays to automate them.

The same logic holds inside ChatGPT Work and Codex, where your allowance covers more tasks on Terra. Handing everyday work to Terra leaves more of your Sol allowance for the hard problems.

The hardest coding still belongs on Sol or Astra

For the most complex coding, pick GPT-6.1 Sol or GPT-6 Astra rather than Terra. OpenAI says so in the launch post, despite Terra's jump to 50.3% on Terminal-Bench 4.0.

A practical pattern is to start coding tasks on Terra and escalate only the ones it gets wrong to Sol (for how Sol and Luna landed in ChatGPT, see GPT-6 in ChatGPT: release dates and Intelligent UI). That costs less than running everything on Sol.

Start on Terra and escalate only the hard jobs

Put GPT-6.1 Terra at max effort and you get roughly Sol's accuracy on OpenAI's factuality test at half Sol's per-token price. Its max-effort error rate of 4.3% is about the same as Sol's, and OpenAI also has it finishing tasks more than three times faster. So when accuracy matters, try raising Terra's effort before you switch models.

If you're still on GPT-5.6 Terra, moving over halves the input price and cuts output by about 58%, and the new model beats the old one on every test OpenAI published. Business plans aren't named in the launch post, so it isn't clear yet whether Terra shows up in their ChatGPT model menu. For a Business team, the reliable option is the API, calling gpt-6.1-terra directly.