GPT-6.1 Terra tested: every check passed, for half a cent a job on low effort

OpenAI released GPT-6.1 Terra on 26 October 2026. It is a new model between the bigger GPT-6.1 Sol and the small GPT-6 Luna, at half the price of Sol. We gave it our secret coding test the next day. Terra passed all 186 checks, on its normal setting and on a lower one. On the lower setting, one coding job took about 12 seconds and cost about half a cent.

What is GPT-6.1 Terra?

GPT-6.1 Terra is OpenAI's newest model. OpenAI's model page calls it "a fast, efficient model for visual reasoning, precise computer use, and clear, natural writing". Computer use means the AI clicks and types on a screen for you.

It costs $1 per million input tokens, $0.05 for input it has seen before (cached), and $5 per million output tokens. That puts it between GPT-6.1 Sol ($2 and $10) and GPT-6 Luna ($0.10 and $0.50).

Terra has six "effort" levels: none, low, medium, high, xhigh and max. Effort sets how long the model thinks before it answers. Medium is the default. You can use Terra in OpenAI's API, in ChatGPT Work and Codex on Plus and higher plans, and in GitHub Copilot. It was not in normal ChatGPT chat when we checked.

Our test

We used our secret coding test, the same one we give every model. It has three jobs: fix the money maths in a small program, write a bill splitter, and write a CSV reader. A CSV reader turns a text file with commas into a table. Each job came in one message, and the model had to answer with the full code. Hidden tests then checked it: 186 checks over three tries per job. More in how we test AI models.

We ran Terra twice through OpenAI's own API, the way developers use it: once on medium effort and once on low. Only Terra got a low-effort run; the other models ran on their default settings. We compare it with earlier runs on the same test: GPT-6.1 Sol, GPT-6 Luna and Claude Haiku 5.5.

The result: Terra passed all 186 checks in both runs. Sol and Haiku also passed all 186. Luna passed 184: it missed two checks on one money-maths try.

What one coding job costs

We worked out the cost from the tokens of each answer. All five match each model's list price:

• GPT-6 Luna: about 0.12 cents per job (184 checks).
• Claude Haiku 5.5: about 0.19 cents (186 checks).
• GPT-6.1 Terra, low effort: about 0.51 cents (186 checks).
• GPT-6.1 Terra, medium effort: about 0.93 cents (186 checks).
• GPT-6.1 Sol: about 1.25 cents (186 checks), on its default setting.

So Terra was cheaper than Sol, but not the cheapest. Haiku and Luna did the same jobs for less. Terra wrote more than Sol per answer, so on medium it was only about a quarter cheaper per job, not half. See our AI API price comparison for more models.

How fast was it?

A typical job took about 21 seconds on medium effort and 12 seconds on low. That is from the question to the full answer.

For comparison: GPT-6.1 Sol took about 22 seconds, Claude Haiku 5.5 about 15 and GPT-6 Luna about 25. So Terra on low was the fastest on average in our test. But the runs were on different days, and Luna and Haiku went through OpenRouter, which adds a small extra step. So take the small gaps lightly. If you need even more speed from Sol, see our Ultrafast test.

Medium or low effort?

On medium effort, Terra thought for about 1,034 tokens per job before it wrote the code. On low, it thought for about 269. That made low almost twice as fast and about half the price, with the same score: 186 of 186.

So for small, clear coding jobs like ours, low effort was enough. For bigger or trickier work, more thinking may help. Our test is too small to show where that line is.

GPT-6.1 Terra in ChatGPT and Codex

You do not need the API to use Terra. On OpenAI's Codex pricing page, it is available in Codex and ChatGPT Work on Plus and higher plans.

On ChatGPT Plus ($20 a month), Codex has a limit per 5 hours. OpenAI estimates these numbers of messages on your own computer (local messages); bigger tasks use more:

• GPT-6.1 Terra: 35 to 285 messages per 5 hours.
• GPT-6.1 Sol: 15 to 160 messages.
• GPT-6 Luna: 350 to 3,000 messages.

So on Plus, Terra gives you roughly twice as many messages as Sol. The Pro plans currently have no 5-hour limit; see our ChatGPT Plus vs Pro test. We did not test Terra in Codex itself.

Should you use it?

• You want the cheapest code that passes? On our test, Claude Haiku 5.5 and GPT-6 Luna were cheaper per job.
• You want an OpenAI model that passed everything, fast? Terra on low effort: 186 of 186 for about half a cent per job.
• You use Codex on ChatGPT Plus? Terra gives you more messages per 5 hours than Sol, by OpenAI's estimate.
• You do big, hard coding work? Our small jobs cannot show whether Sol or Astra does better there.

Terra is a strong, fair-priced model. It just has strong, cheaper rivals.

How we checked

On 27 October 2026 we read OpenAI's model page and API changelog for GPT-6.1 Terra, the Codex pricing page and changelog, and GitHub's changelog.

We ran the second version of our secret coding test through OpenAI's API, the Responses endpoint. It has three jobs, three tries each, and 186 hidden checks, with one request per job. We did this twice: with the default effort (medium) and with effort set to low. We measured the time ourselves. Token counts come from OpenAI's replies, and we worked out the cost from those tokens at OpenAI's list price.

The results for the other models come from our earlier runs of the same test. GPT-6.1 Sol ran through OpenAI on 9 October. GPT-6 Luna and Claude Haiku 5.5 ran through OpenRouter on 8 October; the costs OpenRouter reported match their list prices. The charts are our own, and none of the images was made by AI. Three tries per job is a small test, so your results may differ.