GPT-6.1 Terra:
Fast Computer Use, Lower Cost

01 / Full technical specifications

Important distinction: tool support means a properly configured API workflow can invoke those tools. It does not mean the raw model can independently operate a computer without tool access, permissions and an execution environment. Image generation is a supported tool, not a native image-output modality for Terra.

02 / API pricing: standard rates and long-context adjustments

The following prices are in US dollars per one million text tokens at standard API rates. They are not ChatGPT subscription prices.

Long-context caveat: When an input prompt exceeds 272,000 tokens, OpenAI lists 2× input and cache rates and 1.5× output rates for the entire request, not only the tokens above that threshold. Tool calls may carry separate charges. Thus a 1.05M-token window does not imply that every prompt is billed at the base tier.

Standard token pricing: Terra versus Sol versus Astra

Illustrative totals exclude tool calls, context-size multipliers, cache writes and surcharges. See official API pricing.

At standard rates, Terra is half the token price of Sol and one-tenth of Astra. These are per-token comparisons; total cost per task also depends on reasoning token use, tool calls, execution retries and performance. OpenAI also says Terra's standard API pricing is 50% lower than the previous GPT-5.6 Terra.

03 / Head-to-head: GPT-6.1 Terra vs GPT-6.1 Sol vs GPT-6 Astra

Terra's launch story becomes much more useful when compared with today's two stronger OpenAI alternatives, rather than only its predecessor. All three models have the same advertised context and output limits, but they differ in reasoning options, token rates, typical strengths and success rates on demanding workflows.

Model specifications: same context, different priorities

Official specifications: Terra, Sol, Astra. The later knowledge cutoff of Terra does not imply greater reasoning ability.

Six benchmarks: all three models at maximum reasoning effort

These figures are transcriptions of OpenAI's October 26 launch charts, cross-checked against Kingy AI and MultiAI. These are vendor-reported results, not independent reruns. Scores are percentages. Higher is better except for factual errors.

Real task cost: lower prices do not always win

On Agents' Last Exam v1 at max effort, OpenAI's chart estimates $1.01 per task for Terra, $2.00 for Sol and $7.24 for Astra. OpenAI also says Terra completes the evaluated professional tasks more than three times faster than Sol and Astra. This is measured task completion time, not a universal tokens-per-second advantage.

Yet on OSWorld 2.1 at max effort, the displayed estimated cost is about $1.19 per task for Terra versus $1.15 for Sol. And on Terminal-Bench 4.0, Sol at medium effort reaches 52.7% at about $1.02 per task, beating Terra at max effort (50.3%, approximately $1.84 per task). Both cost examples are reported in the published launch-chart transcription. Token consumption, reasoning depth, tool use, and retries can outweigh list-price savings.

04 / Coding: more than twice the earlier published score

On Terminal-Bench 4.0, an agent coding test performed from the command line, the new model solves 50.3% of tasks at maximum reasoning effort. OpenAI reports 21.5% for GPT-5.6 Terra, a 28.8 percentage-point increase. That is a benchmark improvement, not a claim that all software engineering becomes twice as reliable.

OpenAI nevertheless recommends GPT-6.1 Sol or GPT-6 Astra for the hardest coding tasks. On Terminal-Bench-Science 0.1, which uses code for scientific work, Terra at max reasoning reportedly beats Sol while costing less, and finishes within five points of Claude Opus 5.5 at under one-fifth the estimated cost. These are different benchmarks: the science result does not establish that Terra is the stronger model across general coding.

05 / Computer use and visual reasoning

OpenAI describes Terra as built to understand what's on screen and take actions in applications. Examples include taking measurements from drawings, turning architectural plans into 3D assets, entering invoice data and reviewing video frame-by-frame for annotation workflows.

On MMMU-Pro (Vision), OpenAI reports that Terra surpasses GPT-5.6 Terra at every tested reasoning setting and approaches Astra and Sol in the published visual-reasoning evaluations. Exact numeric chart datapoints are not reproduced here when unavailable in the accessible text of the source; the qualitative comparison is explicitly the manufacturer's claim.

At maximum effort, OpenAI also reports that Terra uses reasoning tokens more than three times as efficiently as Claude Opus 5.5 on Agents' Last Exam. Combining model pricing and evaluated token usage, the company estimates about 9× lower cost per task in that comparison. It is not a universal price comparison for every prompt.

06 / Factual accuracy and safety

On OpenAI's internal, error-inducing question set, factual error rates improved substantially. The measurement matters, but it should not be presented as an externally validated hallucination rate for all user questions.

At max reasoning effort, OpenAI says Terra's error rate on that same evaluation falls to 4.3%, 18 percentage points below its predecessor. The company also reports better adherence to task scope and resistance to prompt-injection attempts in its evaluations.

07 / Availability and how to use it

from openai import OpenAI

client = OpenAI()
response = client.responses.create(
    model="gpt-6.1-terra",
    reasoning={"effort": "medium"},
    input="Summarize the trade-offs of this design."
)
print(response.output_text)

The example assumes a configured OpenAI API key and an installed compatible SDK. It makes a text request only; to operate software or analyze images, add the relevant input and permitted tool configuration.

08 / Should developers choose Terra, Sol or Astra?

Editorial assessment: Terra is compelling when you can verify outputs automatically and run many repeatable tasks. The lower token cost can matter more than small differences in frontier scores. For long, ambiguous or high-stakes workflows, evaluate Sol and Astra on your own examples before optimizing for price. Benchmarks are useful signals, but not substitutes for testing your target workflow.