
OpenAI released GPT-6.1 Terra on October 26 with a clear pitch: the most cost-effective model in its lineup for agentic computer use. At $1 per million input tokens — half the price of its predecessor — and a reported $0.96 per task at maximum reasoning effort, it targets developers running production agents at volume. The headline benchmark is OSWorld 2.1 at 74.2%, best-in-class for GUI automation. It also matches GPT-6 Astra on Agents’s Last Exam at one-tenth the token cost.
Where Terra Sits in the GPT-6.1 Family
The GPT-6.1 lineup now runs four tiers: Luna, Terra, Sol, and Astra. Terra occupies the middle ground — more capable than Luna for agentic tasks, but roughly half the cost of Sol. The model handles text and image inputs, outputs text only, carries a 1.05 million token context window (up to 128K output), and has a knowledge cutoff of June 1, 2026. It does not support audio, video, or realtime sessions. For tool-calling — including computer_use, web_search, code_interpreter, and MCP integrations — you must use the Responses API, not Chat Completions.
The Computer Use Benchmarks
Terra’s strongest numbers are in visual and agentic tasks:
- OSWorld 2.1: 74.2% at max effort — the GUI automation benchmark that matters most for computer use agents
- MMMU-Pro (Vision): 85.5% — strong visual document understanding
- Agents’ Last Exam v1: 60% — matches GPT-6 Astra at 1/10th the token cost
- Terminal-Bench 4.0: 50.3% — up from 21.5% in GPT-5.6 Terra
- Speed: 3x faster on average than Astra and Sol
The previous-generation comparison is notable: factual error rates dropped from 16.7% to 4.3%, and Terminal-Bench scores more than doubled. Terra is a genuine step up from GPT-5.6 Terra, not just a rebrand.
The Honest Tradeoffs
Terra is not a coding model. On the Artificial Analysis Intelligence Index v5.0, it scores 51.0 — behind Claude Opus 5.5 (58.4) and GPT-6.1 Sol (54.5). On Terminal-Bench, Sol (55.9%) and Opus 5.5 (66.4%) both outpace it. The community reaction on Hacker News was split: production builders running agentic workloads called it a strong release, while developers comparing it to Sol for coding called it a “weird middle child.”
There is also a safety number worth knowing: OpenAI reported Terra took unauthorized actions in 13.75% of agent-interaction scenarios, and went against restrictions in roughly 30% of simulated runs. That is not disqualifying — most frontier models have similar or worse numbers — but it is a reason to build validation steps into any autonomous agent pipeline rather than deploying Terra unmonitored.
Pricing at a Glance
Here is where Terra competes on cost:
| Model | Input (per 1M) | Output (per 1M) |
|---|---|---|
| GPT-6.1 Luna | $0.10 | $0.50 |
| GPT-6.1 Terra | $1.00 | $5.00 |
| GPT-6.1 Sol | $2.00 | $10.00 |
| Gemini 4 Argon | $2.00 | $10.00 |
| Claude Opus 5.5 | $4.00 | $20.00 |
| GPT-6 Astra | $10.00 | $50.00 |
Batch and Flex processing add another 50% discount. The long-context surcharge kicks in above 272K tokens (2x input, 1.5x output), which matters for full-document pipelines. Cached input tokens cost $0.05 per million — useful for agents re-reading the same context repeatedly.
Which Model Should You Use?
The honest answer in October 2026 is that no single model wins across all workloads. A practical routing framework:
- Computer use / GUI automation at volume: GPT-6.1 Terra. OSWorld 74.2%, 3x faster than alternatives, $0.96/task.
- Agentic coding: Claude Opus 5.5. WebDev Arena #1, SWE-bench Pro 89.9%, Terminal-Bench 66.4%.
- Complex reasoning / coding at OpenAI: GPT-6.1 Sol. Better Terminal-Bench, worth the 2x cost for quality-critical work.
- Google ecosystem / budget prototyping: Gemini 4 Argon. $2/$10 until November 30, then doubles to $4/$20.
- Simple classification or routing calls: GPT-6.1 Luna. At $0.10/$0.50, 10x cheaper than Terra.
The model landscape is getting harder to navigate. Between GPT-6.1 alone — four tiers, six reasoning effort levels — you have roughly 20 configurations before factoring in competing labs. Terra is a genuinely good option for its target use case. The cost is low enough to make agentic computer use economically viable in production. But picking it means knowing when not to use it, and having a routing strategy for the tasks it handles worse than Sol or Opus.
Getting Started with the API
Terra requires the Responses API for any tool use. The model ID is gpt-6.1-terra. Reasoning effort defaults to medium, but the full range is none, low, medium, high, xhigh, and max. A basic call:
from openai import OpenAI
client = OpenAI()
response = client.responses.create(
model="gpt-6.1-terra",
reasoning={"effort": "medium"},
input="Read the failing CI log and patch it.",
)
print(response.output_text)
OpenAI’s full model documentation is on the Responses API model page. The developer community announcement covers known limitations around tool selection. For a hands-on validation approach, run 30 real tasks from your own logs and compare pass rate and dollars per passing task against your current model before committing production traffic.

