OpenAI's GPT-6.1 Terra cuts API prices and closes in on its bigger models

A cheaper model that scores higher

OpenAI has released GPT-6.1 Terra, a mid-range model in its lineup. As The Decoder reported, the company says it pushes the balance between cost and quality further than any earlier OpenAI model. Terra sits between small, cheap models and large, expensive ones, and OpenAI aims it at everyday office work.

The pricing is for the API, the connection that lets other software use OpenAI's models. Developers pay by the token, a small chunk of text. Terra's input tokens cost $1 per million and its output tokens $5. The previous Terra, GPT-5.6 Terra, charged $2 and $12. Anthropic's Claude Sonnet 5.5 charges $2 and $10.

Cached inputs, meaning text the system has already seen and can reuse, cost just $0.05 per million tokens. That is a quarter of the previous rate.

What the benchmark numbers say

Much of the evidence comes from Artificial Analysis, an outside group that scores AI models. At its highest reasoning effort setting, Terra gains 6 points on the group's Intelligence Index over the older Terra. It costs $0.40 per task instead of $1.40. At lower effort settings, the gains are 7 to 13 points.

Artificial Analysis says this puts Terra on the Pareto frontier. That is the point where you can no longer improve both cost and performance at once. It also calls the model a new high for intelligence per dollar.

On a new Coding Agent Index, Terra scores 48. That is six points below the larger GPT-6.1 Sol at 54, and 17 points above GPT-6 Luna at 31. On Terminal-Bench 4.0, which tests command-line coding tasks, Terra solves 50.3 percent of tasks at maximum effort, up from 21.5 percent for the older Terra. Artificial Analysis says that puts it in the same tier as Sol.

OpenAI itself still recommends GPT-6.1 Sol and GPT-6 Astra for the hardest coding work.

Controlling a computer, and guessing less

OpenAI pitches Terra as strong at operating software the way a person would: clicking through screens, filling out forms, reading technical drawings. On the Agents' Last Exam benchmark, it matches the larger GPT-6 Astra at one-tenth the token price. The company also says it finishes more than three times faster than both Astra and Sol.

Against Claude Opus 5.5 at maximum effort, OpenAI says Terra is more than three times as token-efficient at a quarter of the token price. That works out to roughly nine times lower cost per task. These are the company's own comparisons, so independent testing will matter.

The drop in hallucinations is arguably the biggest change. A hallucination is when an AI makes up an answer instead of admitting it doesn't know. On the private AA-Omniscience test, Terra guessed 38 percent of the time when it lacked the answer, down from 88 percent for the older Terra. Accuracy also rose by two points.

Safety claims and who gets it

Terra meets what OpenAI's Preparedness Framework calls the "Critical" cybersecurity threshold, and it inherits the safeguards of GPT-6 Astra. The framework is the company's own system for rating risks. OpenAI says the new model is more likely than its predecessor to stay within its assigned boundaries and to tell users about its restrictions.

The system card is OpenAI's technical safety document. It says Terra resists direct attacks on the instruction hierarchy about as well as Astra and Sol, which score 99.99 percent on those tests. For indirect prompt injection, where malicious instructions hide in third-party content such as a web page, its defense rate is 98.8 percent. That is up from 96.2 percent for the older Terra.

Terra is available now to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and through the API. The source does not mention any change to subscription prices for ordinary users.

What this means in practice

Developers and software companies are the ones who see the price cut first. If a tool you use runs on OpenAI's models, its maker now pays less to run it. Whether that shows up as lower prices or more features for you is up to each company.

The larger point is that a mid-priced model is now closer to the top tier on several tests. That makes it more practical to have AI handle routine computer tasks, such as drafting a proposal or filling out forms. Treat the benchmark claims with some care, though. Several come from OpenAI, and benchmarks measure specific tests, not your daily work.