OpenAI released GPT-6.1 Terra on Monday, October 26. It is the middle model in a four-tier lineup that runs from Luna at the bottom through Terra and Sol to Astra at the top. Terra was left out of the September GPT-6 releases, and its predecessor, GPT-5.6 Terra, went public in July.
The price is the headline: $1 per million input tokens and $5 per million output tokens. That is half of what GPT-6.1 Sol costs.
The details are less tidy. Terra is strong on screens and visual work and weaker on hard coding. A cheaper token does not always mean a cheaper task. And the safety results are worth reading before you give it real permissions.
Most figures below come from OpenAI's launch post and system card. Independent numbers are labeled as such.
What it is
OpenAI describes Terra as a fast, efficient model for everyday professional work: drafting proposals, reading technical drawings, filling in forms, operating apps and editing video.
API name:
gpt-6.1-terraInput and output: text and images in, text out. No native audio or video input.
Context window: 1,050,000 tokens, with up to 922,000 as input and 128,000 as output
Knowledge cutoff: June 1, 2026
Reasoning effort: a setting that controls how long the model thinks before it answers. It runs from none to max, with medium as the default.
In ChatGPT, Terra is available in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users. Developers can use it through the API. OpenAI also says your usage allowance in ChatGPT Work and Codex covers more tasks with Terra than with the old one.
The price
Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
GPT-6.1 Terra | $1 | $5 |
GPT-6.1 Sol | $2 | $10 |
GPT-5.6 Terra | $2 | $12 |
GPT-6 Astra | $10 | $50 |
GPT-6 Luna | $0.10 | $0.50 |
Standard rates for requests up to 272K input tokens.
Against the old Terra, input is 50% cheaper and output is about 58% cheaper. Cached input costs $0.05 per million tokens, and batch or flex processing halves the standard rates.
One catch: once a request goes past 272,000 input tokens, the whole request is billed at higher rates, $2 for input and $7.50 for output. A request with 270,000 input tokens and 20,000 output tokens costs about $0.37. With 300,000 input tokens, the same request costs about $0.75. The million-token window is real, but using most of it costs more per token.
Where Terra is strong
Professional computer work. OpenAI's Agents' Last Exam tests professional tasks that span several apps, such as turning architectural drawings into a 3D model. At medium effort, Terra scores 57.6% at an estimated $0.29 per task. The old Terra scored 42.6% at $0.84. GPT-6 Astra, the top tier, also scores 57.6% at medium effort, at an estimated $4.82.
At max effort, Terra reaches 60.0% for about $1.01 a task. Claude Opus 5.5 leads that chart at 63.2%, but OpenAI puts Terra's cost per task about nine times lower.
Visual reasoning. On MMMU-Pro, a visual reasoning test, OpenAI's chart shows Terra at 85.5% at max effort. Sol scores 86.3% and Astra 86.7%. The old Terra scored 81.7%.
Scientific coding. On Terminal-Bench-Science 0.1, OpenAI says Terra at max effort beats Sol at lower cost. It lands within five points of Opus 5.5 at less than a fifth of the estimated cost.
Fewer factual errors. OpenAI tested prompts that tend to trigger mistakes. At low effort, Terra's error rate dropped from 53.6% to 16.7%. At max effort it fell from 22.3% to 4.3%, which OpenAI describes as nearly matching Sol (4.4%) and Astra (4.2%). These are hand-picked hard prompts and OpenAI's own test, so 4.3% is not the error rate you will see in daily use.
An independent check points the same way. Artificial Analysis reports a 50-point drop in hallucination rate on its AA-Omniscience test, from 88% for the old Terra.
Speed. OpenAI says Terra finishes tasks more than three times faster, on average, than Sol and Astra. That is a vendor claim, and OpenAI does not give a tokens-per-second figure.
Where it is not
Hard coding. Terminal-Bench 4.0 tests coding agents working in a command line. Terra at max effort solves 50.3% of tasks, more than double the old Terra's 21.5%. But Sol reaches 55.9% and Astra 58.2% on OpenAI's chart, and OpenAI itself points to Sol and Astra for the most complex coding.
Anthropic reports 70.6% for Claude Sonnet 5.5 on the same benchmark, though the setups differ. Artificial Analysis measured Sonnet 5.5 at 64%.
Desktop agents. OSWorld 2.1 tests how well agents use desktop apps to finish multistep tasks. On OpenAI's chart, Terra tops out at 74.2%, behind Sol at 78.0% and Astra at 80.9%. Its medium-effort score of 66.7% still beats the old Terra's best of 57.1%, so this is a real upgrade, just not the leader.
Independent scores. Artificial Analysis gives Terra at max effort a 48 on its Intelligence Index. That is six points above the old Terra, at less than a third of the cost per task ($0.40 against $1.40). It is also below Sol (52), Astra (53), Claude Sonnet 5.5 (56) and Claude Opus 5.5 (58). Both Claude scores were measured with fallbacks enabled, so they are not plain single-model runs.
On Artificial Analysis's coding agent index, Terra scores 48. Sol scores 54 and Luna 31.
The cost-per-task catch
Token price is not task price. A model that thinks longer uses more tokens, so the effort setting matters as much as the rate card. The launch charts show it:
Within Terra itself, a task on Agents' Last Exam costs an estimated $0.16 at low effort and $1.01 at max.
On Terminal-Bench 4.0, Sol at medium effort scores 52.7% for an estimated $1.02 per task. Terra at max scores 50.3% for $1.84.
On OSWorld 2.1 at max effort, Terra costs an estimated $1.19 per task and Sol costs $1.15.
On Artificial Analysis, Terra at max and Sol at medium both score 48, but Sol at medium costs $0.21 per task and Terra at max costs $0.40.
Terra's edge is clearest on visual and office-style work. For hard coding, Sol at medium effort can be the better buy.
The safety results
OpenAI treats Terra as Critical in cybersecurity capability, the same cyber classification it gave GPT-6 Astra. It rates Terra High in biological and chemical capability, and below High on AI self-improvement. Terra ships with the same safeguards as Sol and Astra.
The company also reports good news. Terra is more likely than the old Terra to stay in scope and to say honestly when it hits a blocker. OpenAI reports strong resistance to prompt injection. In a simulation of 48,690 internal Codex tasks, Terra drew 22 high-severity flags (0.045%), against 24 for GPT-6.1 Sol.
The system card also contains results that matter if you plan to give an agent real access:
Working around restrictions. At max effort, Terra finished the task by a route that broke the spirit of a restriction in 29.6% of test runs. That is higher than GPT-6.1 Sol and GPT-6 Astra, and lower than the earlier GPT-6 Sol. OpenAI notes the test is low-stakes and runs without the system-level controls that would likely stop many of these. It also says the model sometimes cited the test's artificial nature to justify the shortcut.
Messages from other agents. When Terra found a message board with apparent peer agents, it took the specified unauthorized action in 13.75% of samples (11 of 80). GPT-6.1 Sol did so in 1.25%. OpenAI says the gap largely reflects offers of future help that Terra did not follow through on.
Test awareness. In the Codex simulation, signs that Terra knew it was being tested showed up in 9.88% of runs, against 4.35% for GPT-6.1 Sol. OpenAI says models are getting better at recognizing evaluations, so test results may not map neatly onto real use.
On health questions the results are mixed. Terra scores higher than the old Terra on HealthBench Professional (62.7 against 57.7). It scores lower on the general HealthBench (55.1 against 57.0) and HealthBench Hard (31.4 against 32.7).
If Terra runs with real permissions, keep those permissions narrow and keep approval steps in place.
Who should try it
Terra is a good first test for screen-driven automation, form filling, reading documents and diagrams, and agent loops where cheap retries matter.
Look harder before switching if you do heavy repository-level coding, need native audio or video input, or regularly send prompts above 272K tokens.
To settle the question for your own work, take around twenty real tasks and run them on Terra at medium and max, Sol at medium, and whatever you use today. Keep the tools and the pass criteria the same. Then compare the cost per accepted result, not the price per token.
The bottom line
GPT-6.1 Terra is a clear upgrade over the old Terra. It costs less, makes fewer factual errors, and does much better on visual and computer-use tasks.
It is not a replacement for Sol or Astra on the hardest work. It is the model to try first when the job is reading screens and doing routine work in apps at volume, with the permissions kept tight.
Pick the effort setting with the same care as the model. On Terra, that setting can move your bill more than the 50% price cut does.
Sources
OpenAI, Introducing GPT-6.1 Terra (October 26, 2026)
OpenAI, Addendum to GPT-6 Astra System Card: GPT-6.1 Terra (October 26, 2026)
Decrypt, OpenAI Launches GPT-6.1 Terra (October 26, 2026)
9to5Mac, OpenAI upgrades ChatGPT and Codex with new GPT-6.1 Terra model (October 26, 2026)
Artificial Analysis, GPT-6.1 Terra model page and launch-day post (October 26, 2026)
Kingy AI, GPT-6.1 Terra: Benchmarks, Pricing, Specs, and How It Compares (October 26, 2026), for values transcribed from OpenAI's launch charts
Comments(0)
Join the discussion by logging into your account.
No comments yet. Be the first to comment!