OpenAI released GPT-6.1 Terra on October 26, US time. The API model name is gpt-6.1-terra. It is the first refresh of the mid-range Terra line since GPT-5.6 Terra in July, and it leans on two things: reading the screen and operating a computer on a person's behalf, and pushing prices down another notch.
Standard API pricing is $1 per million input tokens, $0.05 for cached input and $5 for output. OpenAI says that is 50% cheaper than both GPT-5.6 Terra and GPT-6.1 Sol. Going by the old prices recorded by third parties, GPT-5.6 Terra cost $2 for input and $12 for output, so the cut on the output side is closer to 60%. Sol's output price is $10, so for Sol the cut is exactly half.
The new model is available starting today to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and developers can call it through the API. Free users do not get it this time; their default remains GPT-5.6 Luna.
Computer use is the headline feature
The use cases OpenAI points to are concrete: reading incoming invoices and entering the data into accounting software, reading technical drawings, filling in forms, and editing video. The benchmark it used is Agents' Last Exam, whose tasks include turning architectural drawings into 3D models and reviewing an animated film shot by shot to write timecoded notes for a repair team.
The company's published results are as follows. At medium reasoning the model scores 15 percentage points higher than GPT-5.6 Terra at about one-third of the cost. It matches the flagship GPT-6 Astra at one-tenth of Astra's price per token, and it finishes tasks more than 3 times faster than both Astra and GPT-6.1 Sol. On OSWorld 2.1, medium reasoning already beats the best result posted by GPT-5.6 Terra.
On coding, the highest reasoning setting completed 50.3% of tasks on Terminal-Bench 4.0, against 21.5% for the figure published for the previous generation. OpenAI itself adds that for the most complex coding tasks it still recommends Sol or Astra.
The factual error rate is another major change. On an OpenAI-built set of questions designed to induce mistakes, the error rate at low reasoning fell from 53.6% to 16.7%, and at the highest setting it dropped to 4.3%. On the safety side, under OpenAI's Preparedness Framework, Terra's cybersecurity capability has reached the Critical tier. It reuses the safeguards built for GPT-6 Astra, with details in the system card.
How it stacks up against its peers
All of the figures above come from OpenAI's internal testing. The third-party firm Artificial Analysis also published independent results the same day. The methodology differs, but the direction is the same:
- At the highest setting, its Intelligence Index score is 6 points above GPT-5.6 Terra, 4 points below GPT-6.1 Sol and 10 points above GPT-6 Luna;
- The cost of running the full index comes to about $0.40 per task, versus $1.40 for the previous generation;
- It scores 48 on the Coding Agent Index, against 54 for Sol and 31 for Luna;
- Its hallucination rate on AA-Omniscience fell from 88% to 38%, while accuracy rose by 2 percentage points.
Artificial Analysis concludes that the low and medium reasoning settings sit on the optimal frontier of the intelligence-versus-cost curve, while the other settings fall behind it. Put differently, the best way to use Terra is to run large volumes of everyday tasks at low or medium settings. Saving the highest setting for a few hard problems makes little sense, since Sol is the better pick at that point.
Against competitors, OpenAI benchmarks Terra against Anthropic's Claude Opus 5.5: on Agents' Last Exam, Terra's cost per task at the highest reasoning setting is about one-ninth that of Opus 5.5. Opus 5.5 is Anthropic's flagship and Terra is OpenAI's mid-range model, so the two are not positioned equally. The comparison reads more like a message to enterprise customers that work such as having an AI click a mouse does not call for the most expensive model.
For teams already using Sol in Codex and ChatGPT Work, the most direct change is in usage limits. OpenAI says that under the same usage allowance, Terra can complete more tasks than GPT-5.6 Terra, GPT-6.1 Sol and GPT-6 Astra. The company has not given a multiple for how much more, so users will have to run it for a while to find out.
Sources: OpenAI product announcement and system card, Artificial Analysis independent evaluation, CocoLoop, Phoenix New Media Tech; OpenAI API pricing page used to verify the three prices for input, cached input and output.