One dollar per million input tokens and five per million output: that is the price of GPT-6.1 Terra, the mid-tier model OpenAI released on October 26, 2026. It is exactly half of what Claude Sonnet 5.5 costs, and according to Artificial Analysis it beats its predecessor almost across the board, jumping from 21.5% to 50.3% on Terminal-Bench 4.0.
If you run agents in production, the question is not whether Terra is OpenAI's best model (it isn't), but whether it gets the most useful work done per dollar.
What GPT-6.1 Terra brings
Terra is the middle tier of the GPT-6.1 family, below GPT-6.1 Sol and GPT-6 Astra. It is available in the API as gpt-6.1-terra, and in ChatGPT Work and Codex on the Plus, Pro, Enterprise and Edu plans.
How it compares with GPT-5.6 Terra, using Artificial Analysis figures reported by The Decoder:
| Metric | GPT-5.6 Terra | GPT-6.1 Terra | GPT-6.1 Sol |
|---|---|---|---|
| Artificial Analysis Intelligence Index | 42 | 48 | 52 |
| Coding Agent Index v2.0 | — | 48 | 54 |
| Terminal-Bench 4.0 (max effort) | 21.5% | 50.3% | — |
| Guesses when it doesn't know (AA-Omniscience) | 88% | 38% | — |
| Cost per task on the index (max effort) | $1.40 | $0.40 | — |
The number I care most about is AA-Omniscience: when it lacks the answer, the previous model made one up 88% of the time; Terra does it 38% of the time. For an agent that reads documentation or decides which command to run, that matters more than a couple of accuracy points.
The other headline is computer use. OpenAI pitches it as precise UI control with fast visual reasoning, and on OSWorld 2.1 it beats GPT-5.6 Terra's best medium-effort result at a fraction of the cost. Indirect prompt injection defense goes from 96.2% to 98.8%, which is worth knowing if your agent reads web pages or email.
How to try it
The minimal call with the Python SDK and the Responses API:
from openai import OpenAI
client = OpenAI() # reads OPENAI_API_KEY from the environment
resp = client.responses.create(
model="gpt-6.1-terra",
input="Summarize this postmortem in five lines and suggest two new alerts: ...",
)
print(resp.output_text)If you are on GPT-5.6 Terra today, the switch is the model id. Run your own eval set against both ids before moving traffic: the indexes measure generic tasks, not your prompts.
GPT-6.1 Terra vs Claude Sonnet 5.5 and its predecessor
Claude Sonnet 5.5 is the obvious benchmark: Anthropic's mid-tier model and the one many teams use for coding agents and automation. Prices per million tokens from the official pages:
| GPT-6.1 Terra | GPT-5.6 Terra | Claude Sonnet 5.5 | |
|---|---|---|---|
| Input | $1 | $2 | $2 |
| Cached input | $0.05 | $0.20 | $0.10 (cache hit) |
| Output | $5 | $12 | $10 |
Take an agent that processes 10 million input tokens (uncached) and generates 2 million output tokens a day:
- GPT-6.1 Terra: 10 × 1 + 2 × 5 = $20 a day.
- Claude Sonnet 5.5: 10 × 2 + 2 × 10 = $40 a day.
- GPT-5.6 Terra: 10 × 2 + 2 × 12 = $44 a day.
Caching keeps the same ratio: Terra's cached input is half of Sonnet 5.5's cache-hit price, and Sonnet also charges for cache writes ($2.50 per million for the 5-minute duration).
Where it wins and where it loses
It wins on cost per task. OpenAI says that on Agents' Last Exam it matches GPT-6 Astra at a tenth of the per-token price and, against Claude Opus 5.5 at max effort, is more than three times as token-efficient and comes out roughly nine times cheaper per task. On Terminal-Bench-Science 0.1 it lands within 5 points of Opus 5.5 at under a fifth of the estimated cost. Those are OpenAI's comparisons, but they point the same way as the independent indexes.
It wins on speed. OpenAI puts it at more than three times faster than GPT-6 Astra and GPT-6.1 Sol, which shows up in agent loops with lots of short calls.
It loses at the hard end. OpenAI itself still recommends GPT-6.1 Sol and GPT-6 Astra for the most demanding coding work, and Sol is 6 points ahead on the Coding Agent Index. There is also no published head-to-head with Claude Sonnet 5.5 on the same benchmarks, so against Sonnet the confirmed advantage is price, not quality.
Who should switch
Yes, if you run high-volume agents (incident classification, ticket triage, simple PR reviews, desktop automation) where cost per task drives the decision and a lower guess rate means fewer retries. Also if you already build on the OpenAI API: it is a one-line change.
No, or not yet, if your workload is large refactors or long-running coding agents where you already use Sol, Astra or Opus: the vendor's own advice points elsewhere. And if your team has standardized on Claude, half Sonnet 5.5's price is a reason to run your evals, not to migrate blind.
Sources
- Introducing GPT-6.1 Terra, OpenAI, October 26, 2026.
- GPT-6.1 Terra halves OpenAI's API prices and still gains ground on its larger models, The Decoder, October 26, 2026.
- Claude pricing, Anthropic.