GPT-6.1 Terra is OpenAI's fast, low-cost model, released October 26, 2026. Its API price, $1 input and $5 output per 1M tokens, is half of GPT-6.1 Sol's, and at medium effort it matched GPT-6 Astra's 57.6% on Agents' Last Exam.
It succeeds GPT-5.6 Terra and is tuned for visual reasoning, computer use and writing. You can pick it in ChatGPT Work and Codex, or call it through the API, but it isn't in regular Chat in ChatGPT. OpenAI also says Sol and Astra are still the better choice for the hardest coding work.
Current as of October 28, 2026, based on OpenAI's announcement, API docs, ChatGPT release notes and system card. All scores are OpenAI's published numbers.
Read in other languages: 한국어 · 日本語 · Tiếng Việt
GPT-6.1 Terra at a glance
| Item | GPT-6.1 Terra |
| API model ID | gpt-6.1-terra |
| Price per 1M tokens | $1 input / $5 output |
| Cached input | $0.05 |
| Context window | 1,050,000 tokens |
| Max output | 128,000 tokens |
| Knowledge cutoff | June 1, 2026 |
| Input / output | Text, image / text |
| Reasoning effort | none to max (default medium) |
Where Terra sits in OpenAI's lineup
By API price, from highest to lowest per token, the lineup runs Astra, Sol, Terra and Luna. OpenAI describes Terra as its fast, efficient option, and the GPT-5.6 Terra model page says Terra roughly matches the mini tier of earlier GPT-5 families.
| Model (per 1M tokens) | Input / output | Cached input |
| GPT-6 Astra | $10 / $50 | $1 |
| GPT-6.1 Sol | $2 / $10 | $0.10 |
| GPT-6.1 Terra | $1 / $5 | $0.05 |
| GPT-5.6 Terra | $2 / $12 | $0.20 |
| GPT-6 Luna | $0.10 / $0.50 | $0.01 |
OpenAI pitches Terra at the work that eats up office hours: drafting proposals, reading technical drawings, filling out forms and editing video. Its API example is a high-volume job, reading incoming invoices and entering them into accounting software.
Benchmarks: where Terra leads and where it doesn't
OpenAI ran most models at five reasoning effort levels, low through max, and charted the score against the cost per task.
In the first two line charts, the horizontal axis is cost per task on a log scale and the vertical axis is score. Costs range from cents to several dollars, so each tick on that axis is about 3x the one before it. Each line is one model, and its points run from low on the left to max on the right, so the closer a line sits to the top left, the more score you get per dollar.
Professional work: Agents' Last Exam
This test checks whether a model can finish expert jobs on a computer, such as turning architectural drawings into a 3D model or reviewing an animated film shot by shot to log timecodes for a restoration team. For Claude Opus 5.5, OpenAI published only a max result.
At medium, Terra scored 57.6%, the same as Astra at medium, for $0.29 per task versus Astra's $4.82, about one-seventeenth of the cost. Against GPT-5.6 Terra at medium (42.6%, $0.84), it gained 15 points at about a third of the cost.
At max, Terra's 60.0% tops the best scores of Sol (58.6%) and Astra (59.3%), but Claude Opus 5.5 leads at 63.2%. Terra costs $1.01 per task there, against $8.82 for Opus 5.5. The vertical axis starts at 30%, so the gaps between lines look larger than they are. OpenAI also says Terra finished these tasks over three times faster than Astra and GPT-6.1 Sol, without publishing timing data.
Computer use: OSWorld 2.1
OSWorld 2.1 checks whether a model can complete multi-step tasks in desktop apps, such as building a 3D model from a technical drawing.
At medium alone, Terra (66.7%, $0.50) beat GPT-5.6 Terra's best result (57.1% at max, $4.76). It doesn't catch the bigger models here, though. At max, Terra scored 74.2%, GPT-6.1 Sol 78.0% and Astra 80.9%, and Claude Sonnet 5.5 leads at 83.9%. Terra's cost edge on this test is against Astra and Sonnet 5.5: at max it spends $1.19 per task, about a fifth of Sonnet 5.5's $5.74. Against Sol, Terra is cheaper at the four lower effort levels, but not at max: Sol costs $1.15 there and scores higher.
Visual reasoning: MMMU-Pro
MMMU-Pro asks questions that need reasoning over figures, charts and photos.
Astra, Sol and Terra land within 1.2 points of each other, with Terra at 85.5%. OpenAI says Terra beats GPT-5.6 Terra at every effort level. Even Terra at low (83.7%) scores higher than GPT-5.6 Terra at max (81.7%), and Terra's cost per task is about a fifth of Astra's.
Factual accuracy
OpenAI measured how often an answer contained at least one factual error, using hard prompts picked to provoke mistakes. Treat it as a model-to-model comparison, not the error rate you'd see in everyday use.
OpenAI highlights low: from 53.6% on GPT-5.6 Terra to 16.7% on GPT-6.1 Terra, about 37 points lower. At medium the drop is even larger, from 51.5% to 10.1%. At max, Terra's 4.3% is nearly the same as Astra (4.2%) and GPT-6.1 Sol (4.4%). The system card's separate test, built on real conversations where users had flagged errors, also puts Terra on par with GPT-6.1 Sol.
Coding: Terminal-Bench 4.0
Terminal-Bench 4.0 is an agentic coding test run from the command line. One task, for example, asks the model to fix a web page whose layout shifts while it loads.
Terra's 50.3% is more than double GPT-5.6 Terra's published 21.5%, yet it's near the bottom of this group. Claude Sonnet 5.5 leads at 70.6%, and OpenAI itself says GPT-6.1 Sol and Astra remain the better choice for the most complex coding. Terra's cost per task ($1.84) is close to Sol's ($1.98).
On Terminal-Bench-Science 0.1 (science research in code), Terra at max scored 58.3% ($4.15), just ahead of GPT-6.1 Sol (57.0%, $5.47). Astra leads at 68.1% ($23.80), and Claude Opus 5.5, run with a fallback per OpenAI, scored 63.3% ($23.21).
Pricing: half of Sol, and much cheaper output than GPT-5.6 Terra
Terra's standard API price is exactly half of GPT-6.1 Sol's on input, cached input and output. Against GPT-5.6 Terra, input drops by half and output falls from $12 to $5, about 58% lower, so the longer the output, the bigger the savings.
Half the token price doesn't always mean half the bill per task. At max, Terra's cost per task is within about 10% of Sol's on OSWorld 2.1, Terminal-Bench 4.0 and MMMU-Pro. On Agents' Last Exam it's about half ($1.01 vs. $2.00).
Take an agent that sends 100,000 input tokens per request, 90,000 of them cache hits, and gets 5,000 output tokens back. Once the cache is warm, every 1,000 requests cost this much:
The formula is (10,000 × input price + 90,000 × cached price + 5,000 × output price) ÷ 1,000,000. Per request, Terra comes to $0.0395, half of GPT-6.1 Sol ($0.079) and about 40% of GPT-5.6 Terra ($0.098).
The model page lists a few more pricing rules:
- Cache writes: a 30-minute write costs 1.25x input ($1.25), and a 12-hour write costs 2.5x input ($2.50)
- Long prompts: if input goes over 272K tokens, the whole request is billed at 2x for input and cached input and 1.5x for output
- Processing modes: Batch and Flex are half price ($0.50 / $2.50). Fast mode is double ($2 / $10)
- No Ultrafast: the low-latency Ultrafast mode is available only on GPT-6.1 Sol and Astra
Using GPT-6.1 Terra in ChatGPT Work and Codex
Per OpenAI's ChatGPT release notes, Terra is rolling out in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu. Pick GPT-6.1 Terra from the model list. The release notes and the ChatGPT Learn pages don't list Free or Go.
- Not in regular Chat: in ChatGPT, OpenAI's docs list Terra only in Work and Codex, not in regular Chat
- Plan limits don't change: your overall usage limit stays the same, but the same task uses less of it than on Sol or Astra
- Your default may change: some Plus and Pro users may see GPT-6.1 Terra at Medium instead of GPT-6.1 Sol at Light in Work's default Power presets
- Speed modes: Standard and Fast, where available
- Codex CLI: the ChatGPT Learn models page says Terra needs 0.159.0 or later (run
codex -m gpt-6.1-terra). The changelog separately lists 0.165.0 (October 27) as the release that added Terra to the model catalogs
How much more work fits in the same allowance? OpenAI's ChatGPT pricing docs give estimates:
| Model | Plus, per 5 hours | Credits per 1M (in / out) |
| GPT-6.1 Terra | 35–285 messages | 25 / 125 |
| GPT-6.1 Sol | 15–160 messages | 50 / 250 |
| GPT-6 Astra | 5–45 messages | 250 / 1,250 |
The message counts are estimates that swing widely with the task and settings, and weekly limits may also apply. Business (Standard) gets the same range as Plus, and Pro currently has no 5-hour limit. Terra's credit rates are half of GPT-5.6 Terra's on input (25 vs. 50) and about 42% on output (125 vs. 300).
Using GPT-6.1 Terra in the API
Use the model ID gpt-6.1-terra with the Responses API, Chat Completions or Batch. It supports function calling, structured outputs, streaming and prompt caching. Built-in tools such as computer use (computer_use), web search and code interpreter work only in the Responses API. Fine-tuning and Realtime aren't supported.
Unlike GPT-6.1 Sol, Terra supports reasoning effort none, which suits quick classification and extraction jobs that don't need reasoning. Here's a short Python example that pulls fields from an invoice image with the Responses API.
from openai import OpenAI
client = OpenAI()
response = client.responses.create(
model="gpt-6.1-terra",
# none, low, medium (default), high, xhigh, max
reasoning={"effort": "none"},
input=[{
"role": "user",
"content": [
{"type": "input_text", "text": "Return vendor, total and due date as JSON."},
{"type": "input_image", "image_url": "https://example.com/invoice.png"},
],
}],
)
print(response.output_text)
If none gets fields wrong, try low before moving up to a bigger model.
Rate limits depend on your usage tier (Build, Launch and Grow): 5,000 to 15,000 requests per minute and 1M to 40M tokens per minute.
Safety: rated Critical for cybersecurity
The system card says OpenAI treats Terra as Critical for cybersecurity and High for biology and chemistry under its Preparedness Framework. It uses the same safeguards as GPT-6.1 Sol and Astra.
- Exploits for recent vulnerabilities: on OpenAI's ExploitBench Internal Port test (vulnerabilities disclosed from June to August 2026), Terra achieved arbitrary code execution 24.5% of the time. That's higher than GPT-6.1 Sol (21.5%) and lower than Astra (31.5%)
- Prompt injection: Terra scored 99.99% on the instruction-hierarchy test, and the system card says it holds up about as well as Astra and GPT-6.1 Sol against indirect injection
- Alignment: OpenAI says Terra is more likely than GPT-5.6 Terra to stay within the scope it's given and to tell users plainly about limits or blockers
Using GPT-6.1 Terra from the US: residency and plans
For US API teams, the key question beyond access is where your data lives. OpenAI's data controls guide says GPT-6.1 Sol and GPT-6.1 Terra support only US and EU data residency. The US region offers both regional storage and regional processing, so US data residency works with Terra.
- Setup: create an API project with the US region and send requests to
us.api.openai.com - Cost: OpenAI's docs add a 10% uplift for data-residency endpoints on eligible models released on or after March 5, 2026. Terra came out after that date
- Eligibility: the guide says to contact OpenAI's sales team about eligibility
Readers outside the US should check this before building. The UK, Canada, Australia, India, Japan, Singapore and Korea endpoints are storage-only, and Terra isn't available through them. If you need regional handling, US or EU residency is the only option for Terra.
API usage is priced in US dollars. On the ChatGPT side, Plus costs $20 a month, Pro starts at $100 a month, and Business is $20 per user per month billed annually or $25 billed monthly.
Easy to get wrong
Whether Business gets Terra depends on which page you read. The launch post lists Plus, Pro, Enterprise and Edu. The release notes and the ChatGPT Learn changelog, both from the same day, also include Business.
"50% cheaper" has two baselines. Against GPT-6.1 Sol, every rate is exactly half. Against GPT-5.6 Terra, input is half, but output is about 58% lower.
"On par with Astra" covers one result. Terra matched Astra on Agents' Last Exam at medium. On OSWorld 2.1 and Terminal-Bench 4.0, Astra and GPT-6.1 Sol score higher, and Claude Sonnet 5.5 leads both. The practical split is Terra for cost-sensitive, high-volume work and Sol or Astra for the hardest coding and research.
Summary
- GPT-6.1 Terra, released October 26, 2026, succeeds GPT-5.6 Terra and costs $1 / $5 per 1M tokens in the API, half of GPT-6.1 Sol
- It tied Astra on Agents' Last Exam at
medium, and atmaxits factual-error rate fell to Astra's level - In ChatGPT it's in Work and Codex, not regular Chat; in the API it supports only US and EU data residency. Claude models top some charts, and OpenAI still recommends Sol or Astra for the hardest coding
References
- Introducing GPT-6.1 Terra — OpenAI
- GPT-6.1 Terra model — OpenAI API docs
- GPT-5.6 Terra model — OpenAI API docs
- Pricing — OpenAI API docs
- Data controls in the OpenAI platform — OpenAI API docs
- ChatGPT release notes — OpenAI Help Center
- ChatGPT & Codex changelog — ChatGPT Learn
- Models — ChatGPT Learn
- Pricing — ChatGPT Learn
- GPT-6.1 Terra system card addendum — OpenAI
Related posts
- [AI] OpenAI 12-Hour Prompt Caching: Worth the 2.5x Write Cost? — Our guide to OpenAI's 12-hour prompt caching. Terra's 12-hour cache write is 2.5x input ($2.50), so the same break-even math applies.
'English > AI' 카테고리의 다른 글
| [AI] What Is Genie 4? Now With Sound, Not in Ultra Yet (0) | 2026.10.29 |
|---|---|
| [AI] What Is Gemini 4 Carbon? #1 on AA Index, No Release Date (0) | 2026.10.29 |
| [AI] Claude Prompting Tips: Role Prompts Aren't Enough (0) | 2026.10.27 |
| [AI] OpenAI 12-Hour Prompt Caching: Worth the 2.5x Write Cost? (0) | 2026.10.22 |
| [AI] What Is Claude Duet? US Waitlist, Uses Your Limits (0) | 2026.10.22 |
댓글