GPT-6.1 Terra Pricing and Specs: What Does It Cost?

How much does GPT-6.1 Terra cost?

GPT-6.1 Terra costs $1 per million input tokens and $5 per million output tokens. It accepts text and image input, has a context window a bit over one million tokens, allows up to 128K tokens of output, and has a knowledge cutoff of June 1, 2026. It sits in the middle of OpenAI's GPT-6.1 lineup: more capable than Luna, cheaper than Soul.

What are the GPT-6.1 Terra specs?

The model card gives the following.

The model can see images, so screenshots, diagrams and UI captures work as input. That matters because OpenAI's announcement leans on computer use and agentic actions, plus some visual and coding examples. The post is short, with only a few examples.

The sources don't give rate limits, cached-input pricing, batch discounts or exact context-window numbers beyond "a bit over a million." Check the official pricing page before budgeting around any of those.

How does the pricing work out in practice?

Per-request cost is simple arithmetic. A request with 200K tokens of input and 10K tokens of output costs about $0.20 for input plus $0.05 for output, so roughly $0.25.

The 5x gap between input and output pricing shapes how you should use the model. Long-context jobs where you feed in a lot and get back a little (code review, document Q&A, log analysis) stay cheap. Jobs that generate a lot, such as full-project code generation, cost more per token read.

Reasoning effort also affects spend. On OpenAI reasoning models, thinking tokens generally count toward output, so higher effort settings can raise your bill even when the visible answer is short. Confirm how Terra bills them before assuming a flat cost.

How does Terra compare with Luna and Soul?

OpenAI positions Terra between two siblings.

• Luna is cheaper than Terra. Luna did not appear on the benchmark chart in the announcement, and that omission suggests it is a clear step down in capability.
• Soul is the stronger model, especially for coding. On the coding chart, Terra at various effort levels reaches into the territory of Soul's results, but Soul still looks better. Terra is the cheaper trade.

So the family reads as a ladder. Pick Luna when cost is the only constraint. Pick Soul when you need the best coding results. Pick Terra when you want strong results at $1/$5 and care about turnaround time.

One more thing to know when reading the charts: the factual error rate chart on difficult prompts is a lower-is-better metric. Check the axis direction before drawing conclusions from the bars.

What reasoning effort levels does Terra offer?

The announcement shows results at multiple effort levels, and the default is medium. The coding chart cites 50.3% at max effort and 47% at low. That is a gap of about 3 points between the two ends, so on that benchmark, going from low to max buys a modest improvement.

That spread is worth testing on your own workload. If low effort is close enough for your task, you save latency and probably tokens. If you are doing hard multi-step work, max effort is available.

What are fast mode and ultra mode?

Beyond reasoning effort, Terra has two operating modes that a reviewer used in testing.

• Fast mode prioritizes speed. The sources don't say whether it changes pricing, so check that before using it at scale.
• Ultra mode spawned several sub-agents in testing, which split a job across parallel workers.

A reviewer combined these with max effort in different runs. Some timings from that hands-on testing, all single runs with no controlled comparison:

These are anecdotes, not benchmarks. They do show the pattern: simple outputs arrive in minutes, and fast plus ultra shortens the wait noticeably. In the wrestling run, the model opened and test-played the Godot project itself.

How good are the results?

The reviewer described the output as competent but simple. The skate game had inverted controls on the first pass, and the wrestling game's characters initially faced backwards. Both were fixed with a follow-up polish prompt. The reviewer felt the style resembled a distilled version of a larger sibling, and noted it did less extra detail work, such as pedestrian interaction, than some competing models tend to add.

The odd result was a robot arm control test. Terra asked for a better camera angle, said the servo mapping looked wrong, and eventually stopped without completing the task. The reviewer called it hesitant. That is a single test, but it is a reminder that computer-use and embodied tasks need your own evaluation, even though computer use is the headline pitch.

Is GPT-6.1 Terra worth it?

For many API workloads, the price makes it an easy model to try. At $1 in and $5 out, with a 1M+ context window and 128K output, it covers long documents, large codebases and long generations without forcing you into the top tier.

It is a weaker fit if you need the best coding quality available, since Soul holds that spot. It is also not the choice if you need the lowest possible per-token cost, since Luna is cheaper. Terra's case rests on being the capable-but-affordable middle, plus its speed.