OpenAI's half-price GPT-6.1 Terra only partly delivers, an independent test finds

Filling out a form, editing a video, reading a technical drawing, automating a repetitive task on a computer: that is the job OpenAI has set for GPT-6.1 Terra, launched on October 26. The model replaces GPT-5.6 Terra. In the application programming interface (API), it costs half as much, according to the company.

OpenAI's lineup has four classes of models, from smallest to most powerful: Luna, Terra, Sol and Astra. Terra sits on the second rung. GPT-6.1 Terra is available to Plus, Pro, Enterprise and Edu subscribers in ChatGPT Work and in Codex, OpenAI's coding tool. Developers can access it as gpt-6.1-terra.

What OpenAI claims

The company's promises are ambitious. According to OpenAI, GPT-6.1 Terra matches GPT-6 Astra, its top-end model, at one-tenth of the price per token (the unit of text that is billed). It would also complete tasks more than three times faster than Astra and GPT-6.1 Sol.

On an internal test made of trick questions, its error rate at low reasoning effort would fall from 53.6% to 16.7% compared with its predecessor. On Terminal-Bench 4.0, a coding test, it would reach 50.3% versus 21.5%, at maximum reasoning effort.

These figures were measured by OpenAI itself, and the company sets a limit: for the most complex coding tasks, it still recommends Sol or Astra.

What an independent evaluator says

Artificial Analysis, which tests AI models independently, published its measurements on launch day. They confirm real progress, especially on cost. At the maximum reasoning setting, GPT-6.1 Terra gains 6 points on its intelligence index over GPT-5.6 Terra, at a cost per task of $0.40 instead of $1.40. It remains 4 points below GPT-6.1 Sol and 10 above GPT-6 Luna. At lower reasoning settings, the gains are clearer: 7 to 13 points.

The most striking change concerns made-up answers. Artificial Analysis measures this with AA-Omniscience, a test of its own. When GPT-6.1 Terra doesn't know the answer, it gives a random one in 38% of cases, versus 88% for its predecessor. The rate of correct answers also improves, by 2 points. That is a clear improvement, but it still means more than one time in three. On most other tests, the improvement is more modest.

This evaluation does not verify the parity with Astra that OpenAI announced. Artificial Analysis ranks Terra below Sol, a model that OpenAI itself positions below Astra.

Who is it for, and at what price?

In the API, GPT-6.1 Terra costs $1 per million input tokens and $5 for output, versus $2 and $12 for GPT-5.6 Terra, according to Artificial Analysis. For a company running AI agents on repetitive tasks, the drop is significant. However, the sources consulted do not say whether the model is offered to free users or to ChatGPT Business accounts.

As with Kimi K3.1 Pro, another model ranked by Artificial Analysis, a good score is no substitute for testing the model on your own use cases before switching, especially for tasks where a made-up answer is costly. Other independent evaluations are expected in the coming weeks.